Autonomous Cooperation and Control in Logistics
Michael Hülsmann Katja Windt
Bernd Scholz-Reiter
Editors
Autonomous Cooperation and Control in Logistics Contributions and Limitations Theoretical and Practical Perspectives
123
Editors Michael Hülsmann Jacobs University Bremen Systems Management - International Logis Campus Ring 1 28759 Bremen Germany
[email protected] Katja Windt Jacobs University Bremen Global Production Logistics - Internatio Campus Ring 1 28759 Bremen Germany
[email protected] Prof.Dr. Bernd Scholz-Reiter Universität Bremen BIBA GmbH Bremer Institut für Messtechnik Automatisierung und Qualitätswiss. Hochschulring 20 28359 Bremen Germany
[email protected] ISBN 978-3-642-19468-9 e-ISBN 978-3-642-19469-6 DOI 10.1007/978-3-642-19469-6 Springer Heidelberg Dordrecht London New York Library of Congress Control Number: 2011929775 c Springer-Verlag Berlin Heidelberg 2011 This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Cover design: eStudio Calamar S.L. Printed on acid-free paper Springer is part of Springer Science+Business Media (www.springer.com)
Preface
The edited volume “Autonomous Cooperation and Control in Logistics: Contributions and Limitations – Theoretical and Practical Perspectives” consequently continues the previous publication “Understanding Autonomous Cooperation and Control in Logistics – The Impact of Autonomy on Management, Information, Communication, and Material Flow”, edited by Michael Hülsmann and Katja Windt. This first volume focuses on collating various understandings of self-organisation. It intends to establish a common perspective on basic ideas and their adoption and adaptability to logistics. Additionally, that publication identifies and compares the scope and depth of autonomous cooperation and control resulting from various, interdisciplinary understandings of self-organisation. Therefore, the first edited volume aims for developing a conceptual system for autonomous cooperation and control, which allows to interpret discipline-specifically, to functionalise, and to apply autonomous cooperation and control in the context of logistic systems. It is dedicated to provide logistics research as well as practice with first impulses for answering the question how logistics management can better cope with complexity and dynamics in supply chains and networks. All in all, the first edited volume “Understanding Autonomous Cooperation and Control in Logistics – The Impact of Autonomy on Management, Information, Communication, and Material Flow” provides a terminological, taxonomical, and analytical framework to examine, explain, and apply the principles of self-organisation in the context of complex, dynamic logistics processes. Consistently, the second edited volume “Autonomous Cooperation and Control in Logistics: Contributions and Limitations – Theoretical and Practical Perspectives” uses the developed framework to approach the challenge of finding an optimal degree of autonomous cooperation and control of logistics processes. Therefore, this publication seeks to determine analytically the upper and lower boundaries of autonomous cooperation and control. This focus should lead to a common understanding of the enablers and impediments of applying the idea of self-organisation as a paradigmatic principle to logistics and the design, planning, and control of its processes. Hence, this edited volume is dedicated to identify, describe, and investigate – in the context of production and distribution logistics – the effects, feasibility, outcomes, barriers, driving forces, cause-effect-relations, etc. of concepts, methods, technologies, and routines, that are based on and linked with the idea of selforganisation in logistics. Therefore, it is the major objective of this edited volume v
vi
Preface
to give a broad overview about the contributions and limitations of autonomous cooperation and control of logistics processes. Furthermore, this publication should outline a first answer to how the abstract idea of self-organisation can substantially contribute to a better performance and robustness of complex and dynamic logistics systems in versatile and volatile environments. In this regard this edited volume comprises first implementations in real industrial scenarios as well and demonstrates on practical examples the logistic potential and its limitations. Both research objects – logistics as well as the idea of autonomous cooperation and control – need an interdisciplinary approach, which can cover all their heterogeneous characteristics (e.g. technological and organisational). Therefore, this edited volume combines the different perspectives of production technology, electronics and communication engineering, informatics and mathematics, as well as management sciences and economics. The solid foundation for the necessary integration of these diverse points of view is built on the interdisciplinary research within the Collaborative Research Centre 637 (CRC 637) “Autonomous Cooperating Logistic Processes – A Paradigm Shift and its Limitations” at the University of Bremen since 2004. The CRC 637 intends to identify the rules of autonomous cooperation and control in order to develop a “theoretical backbone” for applying this paradigm on all levels of logistics systems: on the managerial decision-making level, on the information processing and communication level, and on the material flow level. Therefore, this publication edited volume covers all perspectives and levels addressed above in order to provide a comprehensive and profound picture of contributions and restrictions of autonomous cooperation and control of logistics processes – which might help to understand the related paradigm-shift and its limitations. This publication is the result of a fruitful and pleasant cooperation, collaboration, and communication between many actors, whose invaluable work made this edited volume possible. First of all, we like to thank our colleagues and doctoral students within the CRC 637 and around this institution for the inspiring, intriguing discourses, reflections, and exchanges of ideas within the last seven years. During our debates and conversations we had the outstanding opportunity to learn from other disciplines. This included also the challenge to develop shared perspectives on the same object (i.e. autonomous cooperation and control in logistics) from the background of different scientific cultures, theoretical frameworks, and methodological approaches. Therefore, it was always an honour and contentment for us having the chance to edit this volume and we are very grateful for this exciting experience in our academic career. Secondly, we are tremendously happy and grateful for the contributions of the voluntary reviewers, who spent their limited and valuable time for improving the quality of the contributions in this edited volume. Without any doubt, the reviewers’ comments formed the various collections of a good idea, an appropriate research conception, and all the other ingredients of a scientific article into the shape of consistent and solid argumentation. There are also very helpful and important hands, which backed us up by thoroughly taking care of all the supporting activities. For this, we would like to express our appreciation to Dipl.-Wi.-Ing. Anne Schwientek, who coordinated the compilation and
Preface
vii
editing of all articles; to Dipl.-Oec. Philip Cordes for supporting and reviewing; to Lore Zander and Dipl.-Inf. Jacub Piotrowski for handling the administrative side; to Susanne Benner and Caroline Hannemann for proof-reading and editing. Naturally, we would like to say thank you to our publisher SpringerPhysica, represented by Thomas Lehnert, for his motivating feedbacks and for giving us the chance to publish our edited volume “Autonomous Cooperation and Control in Logistics: Contributions and Limitations – Theoretical and Practical Perspectives” at SpringerPhysica’s. Finally, we would like to thank the German Research Foundation (DFG), which supported this research as part of the Collaborative Research Centre 637 “Autonomous Cooperating Logistic Processes – A Paradigm Shift and its Limitations”. Michael Hülsmann Bernd Scholz-Reiter Katja Windt
Acknowledgements
This research was supported by the German Research Foundation (DFG) as part of the Collaborative Research Centre 637 “Autonomous Cooperating Logistic Processes – A Paradigm Shift and its Limitations” at the University of Bremen and the Jacobs University Bremen.
ix
Contents
1
Contributions and Limitations of Autonomous Cooperation and Control in Logistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . Michael Hülsmann, Bernd Scholz-Reiter, and Katja Windt
1
Part I Organizational Contributions and Limitations 2
Approaches to Organizational Contributions and Limitations of Autonomous Cooperation and Control in Logistics . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . 11 Michael Hülsmann, Anne Schwientek, and Philip Cordes
3
Creating Customer Value in Logistics: Contributions and Limitations of Autonomous Cooperation-Based Technologies . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . 15 Michael Hülsmann, Anne Schwientek, Benjamin Korsmeier, and Linda Austerschulte
4
Autonomous and Central Control of Production Networks . . . . . .. . . . . . . 27 Sergey Dashkovskiy, Andrii Mironchenko, and Lars Naujok
5
Approaching the Application Borders of Network Capacity Control in Road Haulage .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . 45 Jörn Schönberger and Herbert Kopfer
6
Limit and Degree of Autonomy in Groupage Systems: Challenges, Chances and Barriers for Horizontal Cooperation in Operational Transportation Planning . . . . . . . . . . . .. . . . . . . 61 Heiko Wieland Kopfer, Herbert Kopfer, and Xin Wang
7
The Interaction Effort in Autonomous Logistics Processes: Potential and Limitations for Cooperation . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . 77 Arne Schuldt, Jan Ole Berndt, and Otthein Herzog
xi
xii
Contents
Part II Methodical Contributions and Limitations 8
Introduction to Methodical Contributions and Limitations of Autonomous Cooperation and Control in Logistics . . . . . . . . . . . .. . . . . . . 93 Till Becker and Katja Windt
9
Logistic Systems with Multiple Autonomous Control Strategies . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . 97 Bernd Scholz-Reiter, Michael Görges, and Thomas Jagalski
10 Communities of Autonomous Units: An Approach to Interactive Computation, Its Power and Limitations. . . . . . . . . . .. . . . . . .113 Hans-Jörg Kreowski, Sabine Kuske, Melanie Luderer, and Caroline von Totth 11 Potentials and Limitations of Autonomously Controlled Production Systems. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .131 Bernd Scholz-Reiter, Michael Görges, and Henning Rekersbrink 12 Scalability Effects in Modeling Autonomously Controlled Logistic Processes: Challenges and Solutions in Business Process Modeling . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .149 Bernd Scholz-Reiter, Daniel Rippel, and Steffen Sowade 13 Exploitation of Manufacturing Flexibilities in Decision Methods for Autonomous Control of Production Processes: Findings from Industrial Practice and Theoretical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .169 Nicolas Gebhardt, Oliver Jeken, and Katja Windt Part III
Technological Contributions and Limitations
14 Views on Technological Contributions and Limitations of Autonomous Cooperation and Control in Logistics . . . . . . . . . . . .. . . . . . .191 Jakub Piotrowski and Bernd Scholz-Reiter 15 Implications of Communication Constraints for the DLRP in Transport Logistics .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .195 Bernd-Ludwig Wenning, Christian Zabel, Henning Rekersbrink, Carmelita Görg, and Bernd Scholz-Reiter 16 Embedded Intelligent Objects in Food Logistics Technical Limits of Local Decision Making . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .207 Reiner Jedermann, Javier Palafox-Albarran, Amir Jabarri, and Walter Lang
Contents
xiii
17 Knowledge Management for Agent-Based Control Under Temporal Bounds . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .229 Tobias Warden, Robert Porzel, Jan D. Gehrke, Hagen Langer, Otthein Herzog, and Rainer Malaka 18 Impacts of Data Integration Approaches on the Limitations of Autonomous Cooperating Logistics Processes .. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .247 Karl A. Hribernik, Christoph Kramer, Carl Hans, and Klaus-Dieter Thoben Part IV
Practical Contributions and Limitations
19 Considerations of Practical Contributions and Limitations of Autonomous Cooperation and Control in Logistics . . . . . . . . . . . .. . . . . . .271 Michael Hülsmann, Philip Cordes, and Anne Schwientek 20 A Comparative View on Existing Autonomous Control Approaches: Observations from a Simulation Study . . . . . . . . . . . . . .. . . . . . .275 Till Becker and Katja Windt 21 Limitations of Autonomous Control in Practical Applications: Report on Lessons Learned from Vehicle and Apparel Logistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .291 Bernd Scholz-Reiter, Carmen Ruthenbeck, Michael Teucke, and Jantje Hoppert 22 Autonomous Control in Production Planning and Control: How to Integrate Autonomous Control into Existing Production Planning and Control Structures . . . . . . . . . . . . . . . . . . . . . .. . . . . . .313 Marius Veigt, Farideh Ganji, Ernesto Morales Kluge, and Bernd Scholz-Reiter Index . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .331
Contributors
Linda Austerschulte Systems Management, International Logistics, School of Engineering and Science, Jacobs University Bremen, Bremen, Germany Till Becker Global Production Logistics, International Logistics, School of Engineering and Science, Jacobs University Bremen, Bremen, Germany,
[email protected] Jan Ole Berndt TZI IS, Universität Bremen, Am Fallturm 1, 28359 Bremen, Germany,
[email protected] Philip Cordes Systems Management, International Logistics, School of Engineering and Science, Jacobs University Bremen, Germany,
[email protected] Sergey Dashkovskiy ZeTeM, Universität Bremen, Bibliothekstr.1, 28359 Bremen, Germany,
[email protected] Farideh Ganji Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany Nicolas Gebhardt Global Production Logistics, International Logistics, School of Engineering and Science, Jacobs University Bremen, Bremen, Germany,
[email protected] Jan D. Gehrke TZI – Center for Computing Technologies, University of Bremen, Bremen, Germany Carmelita Görg Communication Networks, University of Bremen, Bremen, Germany Michael Görges Department of Planning and Control of Production Systems, BIBA, University of Bremen, Germany Carl Hans Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany Otthein Herzog TZI – Center for Computing Technologies, University of Bremen, Bremen, Germany,
[email protected] xv
xvi
Contributors
Jantje Hoppert Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany Karl A. Hribernik Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany,
[email protected] Michael Hülsmann Systems Management, International Logistics, School of Engineering and Science, Jacobs University Bremen, Campus Ring 1, 28759 Bremen, Germany,
[email protected] Amir Jabarri Microsystems Center Bremen, University of Bremen, Bremen, Germany Thomas Jagalski Department of Planning and Control of Production Systems, BIBA, University of Bremen, Germany Reiner Jedermann Microsystems Center Bremen, University of Bremen, Bremen, Germany,
[email protected] Oliver Jeken Global Production Logistics, International Logistics, School of Engineering and Science, Jacobs University Bremen, Bremen, Germany Ernesto Morales Kluge Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany Heiko Wieland Kopfer Institute of Machine Elements, Fastening Systems and Product Innovation, University of Siegen, Siegen, Germany,
[email protected] Herbert Kopfer Chair of Logistics, University of Bremen, Bremen, Germany Benjamin Korsmeier Systems Management, International Logistics, School of Engineering and Science, Jacobs University Bremen, Bremen, Germany Christoph Kramer Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany Hans-Jörg Kreowski Department of Computer Science, University of Bremen, Bremen, Germany,
[email protected] Sabine Kuske Department of Computer Science, University of Bremen, Bremen, Germany Walter Lang Microsystems Center Bremen, University of Bremen, Bremen, Germany Hagen Langer TZI – Center for Computing Technologies, University of Bremen, Bremen, Germany Melanie Luderer Department of Computer Science, University of Bremen, Bremen, Germany Rainer Malaka TZI – Center for Computing Technologies, University of Bremen, Bremen, Germany
Contributors
xvii
Andrii Mironchenko ZeTeM, Universität Bremen, Bibliothekstr.1, 28359 Bremen, Germany,
[email protected] Lars Naujok ZeTeM, Universität Bremen, Bibliothekstr.1, 28359 Bremen, Germany,
[email protected] Javier Palafox-Albarran Microsystems Center Bremen, University of Bremen, Bremen, Germany Jakub Piotrowski Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany,
[email protected] Robert Porzel TZI – Center for Computing Technologies, University of Bremen, Bremen, Germany Henning Rekersbrink Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany Daniel Rippel Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany Carmen Ruthenbeck Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany Bernd Scholz-Reiter Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany,
[email protected] Jörn Schönberger Chair of Logistics, University of Bremen, Bremen, Germany,
[email protected] Arne Schuldt TZI IS, Universität Bremen, Am Fallturm 1, 28359 Bremen, Germany,
[email protected] Anne Schwientek Systems Management, International Logistics, School of Engineering and Science, Jacobs University Bremen, Bremen, Germany,
[email protected] Steffen Sowade Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany Michael Teucke Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany Klaus-Dieter Thoben Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany Marius Veigt Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany,
[email protected] Caroline von Totth Department of Computer Science, University of Bremen, Bremen, Germany Xin Wang Chair of Logistics, University of Bremen, Bremen, Germany
xviii
Contributors
Tobias Warden TZI – Center for Computing Technologies, University of Bremen, Bremen, Germany,
[email protected] Bernd-Ludwig Wenning Communication Networks, University of Bremen, Bremen, Germany,
[email protected] Katja Windt Global Production Logistics, International Logistics, School of Engineering and Science, Jacobs University Bremen, Bremen, Germany,
[email protected] Christian Zabel Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany
Associated Editors
Enzo Morosini Frazzon Department of Production and Systems Engineering, Technology Center - CTC, Federal University of Santa Catarina, Caixa Postal 476 - Campus Universitário UFSC, Trindade, CEP 88040-970, Florianópolis, Santa Catarina, Brazil David Gouyon Centre de Recherche en Automatique de Nancy, Faculté des Sciences et Techniques, Université Henri Poincaré, BP 70239, 54506 Vandoeuvre les Nancy Cedex, France Norbert Gronau Lehrstuhl für Wirtschaftsinformatik und Electronic Government, Universität Potsdam, August-Bebel-Str. 89, 14482 Potsdam, Germany Hans-Dietrich Haasis Lehrstuhl für Allgemeine Betriebswirtschaftslehre und Industriebetriebslehre, Fachbereich Wirtschaftswissenschaft, Universität Bremen, Wilhelm-Herbst-Straße 12, 28359 Bremen, Germany Hamid Reza Karimi Faculty of Engineering and Science, University of Agder, Postboks 509, 4898, Grimstad, Norway Kap Hwan Kim Department of Industrial Engineering, Pusan National University, Changjeon-dong, Kumjeong-ku, Pusan 609-735, Korea Simone Kirpal Zentrum für Sozialpolitik – Geschlechterpolitik im Wohlfahrtsstaat – UNICOM-Gebäude, Universität Bremen, Mary-Somerville-Straße 5, 28359 Bremen, Germany Dirk Christian Mattfeld Lehrstuhl Decision Support, Institut für Wirtschaftsinformatik, Technische Universität Braunschweig, Mühlenpfordstraße 23, 38106 Braunschweig, Germany William McKelvey UCLA Anderson School of Management, 110 Westwood Plaza, Los Angeles, CA 90095, USA Nariaki Nishino Department of Technology Management for Innovation, School of Engineering, The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8656, Japan Ranjit Perera Department of Electrical Engineering, University of Moratuwa, Katubedda, Moratuwa, Sri Lanka xix
xx
Associated Editors
Mykhaylo Postan Lehrstuhl für Management und Marketing, National Odessa Maritime University, Mechnikov Str. 34, 65029 Odessa, Ukraine Christian Prins Laboratoire d’Optimisation des Systèmes Industriels, Institut Charles Delaunay, Université de technologie de Troyes, 12, Rue Marie Curie, BP 2060, 10010 Troyes Cedex, France Johannes Schöning Innovative Retail Laboratory, DFKI GmbH Campus D3_2, Stuhlsatzenhausweg 3, 66123 Saarbrücken, Germany Jens Schumacher Wirtschaftsinformatik, FH Vorarlberg, Hochschulstraße 1, 6850 Dornbirn, Austria Gabriele Taentzer Fachbereich Mathematik und Informatik, Philipps-Universität Marburg, Hans-Meerwein-Straße, 35032 Marburg, Germany Ingo J. Timm Wirtschaftsinformatik I, Universität Trier, Universitätsring 15 54296 Trier, Germany Ipke Wachsmuth AG Wissensbasierte Systeme, Technische Fakultät, Universität Bielefeld, Universitätsstr. 25, 33615 Bielefeld, Germany Hans-Peter Wiendahl Institut für Fabrikanlagen und Logistik, Leibniz Universität Hannover An der Universität 2, 30823 Garbsen, Germany
Authors
Austerschulte, Linda Dipl.-Wi.-Ing. Linda Austerschulte studied Industrial Engineering and Management at the University of Bremen. She received her diploma in 2006. From 2006 to 2009, she worked as a research associate and PhD student at the University of Bremen in the working group “Sustainable Systems Development.” Since then, Linda Austerschulte is a research associate at Jacobs University Bremen in the working group “Systems Management” of Prof. Dr. M. Hülsmann. Her research interests include Strategic Management of Organizational Capabilities and the Measurement of Intangible Assets.
Becker, Till Till Becker studied Information Systems at the University of Münster, Germany. After graduation, he worked for 2 years as an IT auditor and consultant before he returned to academia. He is now a member of the Global Production Logistics workgroup of Prof. Dr.-Ing. Katja Windt at Jacobs University Bremen. As a research associate and PhD student, he focuses his research on graph representations of logistics networks and computer simulations.
Berndt, Jan Ole Jan Ole Berndt received his Diploma in Computer Science from the Universität Bremen in 2009. He is a member of the Artificial Intelligence research group of Prof. Otthein Herzog at the Universität Bremen, holding a scholarship granted by the International Graduate School for Dynamics in Logistics.
xxi
xxii
Authors
Cordes, Philip Dipl.-Economist Philip Cordes completed his Diploma in Business Sciences at the University of Bremen in 2008. He is a research associate and PhD student of Prof. Dr. Michael Hülsmann in the department “Systems Management” in the School of Engineering and Science at the Jacobs University Bremen. Since May 2008, Philip Cordes is working for the Collaborative Research Centre 637 “Autonomous Cooperating Logistics Processes – A Paradigm Shift and Its Limitations” funded by the German research foundation (DFG).
Dashkovskiy, Sergey Sergey Dashkovskiy received the MSc degree in Applied Mathematics and Mechanics from the Lomonosov Moscow State University, Moscow, Russia, in 1996 and the PhD degree in Mathematics from the University of Jena, Jena, Germany, in 2002. He is the head of the research group Mathematical Modeling of Complex Systems at the Center of Industrial Mathematics, University of Bremen, Germany. His research interests are in the field of nonlinear control theory and partial differential equations.
Ganji, Farideh Dipl. -Ing. Farideh Ganji works as a technical assistant at the BIBA – Bremer Institut für Produktion und Logistik GmbH in the University of Bremen in the division Intelligent Production and Logistic Systems.
Gebhardt, Nicolas Nicolas Gebhardt studied Mechanical Engineering at Hamburg University of Technology (TUHH). After his studies, he joined the work group of Prof. Dr.-Ing. Katja Windt for Global Production Logistics at Jacobs University Bremen as a research associate. His main research interest is logistics-oriented design and autonomous control.
Gehrke, Jan Jan D. Gehrke received his diploma degree in Computer Science from the University of Bremen in 2005 with a thesis on knowledge-based scene analysis for intelligent vehicles. He joined the research group of Otthein Herzog as a research assistant in 2005 and is since then affiliated to the CRC 637. His research focuses on intelligent agents in logistics as well as knowledge representation and management in MAS.
Authors
xxiii
Görg, Carmelita Prof. Dr. rer. nat. habil. Carmelita Görg is the head of the Communication Networks group (ComNets) at the University of Bremen within the TZI (Center for Computer Science and Information Technology). She is a member of the board of the ITG (Information Technology Society, Germany) and speaker of the ITG working group 5.2.1 on “System Architecture and Traffic Engineering.”
Görges, Michael Dipl.-Wi.-Ing. Michael Görges works as a research scientist at the BIBA – Bremer Institut für Produktion und Logistik GmbH in the University of Bremen in the division Intelligent Production and Logistic Systems.
Hans, Carl Dr.-Ing. Carl Hans is the head of the department “Intelligent ICT for Co-operative Production” at BIBA. As a research scientist, he focuses on the development and integration of ICT solutions in industrial and research projects in areas including Computer Simulation, Artificial Intelligence, and Knowledge Management.
Herzog, Otthein Dr. Otthein Herzog, professor em. of Artificial Intelligence at the Universität Bremen, and professor of Visual Information Technologies at Jacobs University Bremen. His research interests are: Multiagent-Systems, Knowledge Management, Wearable Computing, and Content-Based Multimedia Analysis. He (co-)authored 190C scientific papers and is a member of the German Academy of Science and Engineering.
Hoppert, Jantje M.A. Kult. Jantje Hoppert is a member of the CRC 637 at the University of Bremen. She included the organizational perspective on limitations of autonomous control into the investigated projects.
xxiv
Authors
Hribernik, Karl Dipl.-Inform. Karl A. Hribernik studied Computer Science at the University of Bremen. He joined the Bremen Institute of Industrial Technology and Applied Work Science (BIBA) as a research scientist in 2002. His research focus is on the integration of heterogeneous data sources into item-level Product Lifecycle Management systems. He is currently involved in the subproject "Integration of Logistics Data to support Autonomous Cooperating Logistics Processes” in the CRC637 "Autonomous Cooperating Logistics Processes – A Paradigm Shift and Its Limitations."
Hülsmann, Michael Prof. Dr. Michael Hülsmann is an associate professor of “Systems Management” in the International Logistics department at Jacobs University Bremen. He is head of the subproject “Monitoring of Autonomous Systems” and member of the board within the CRC 637. His main research interests cover various fields of strategic management for logistics and technology-driven value networks. Especially, his research focus is on organizational competences and change, business models, interorganizational cooperation/coordination/communication, and strategic controlling of performance/risk/sustainability.
Jabarri, Amir Amir Jabbari received his MSc in Mechatronics (2006) and BSc in Electrical engineering (2004). He received his PhD from University of Bremen in August 2009 for his research on “Application of Autonomous Fault detection and Isolation in Measurement Systems.” His main research activities are in Artificial Intelligence, Control Systems, and Data Fusion in Wireless Sensor Networks.
Jagalski, Thomas Thomas Jagalski, MSc (Economics and Management), works as a research scientist at the BIBA – Bremer Institut für Produktion und Logistik GmbH at the University of Bremen in the division “Intelligent Production and Logistic Systems.”
Jedermann, Reiner Reiner Jedermann finished his Diploma in Electrical Engineering in 1990 at the University of Bremen. After two employments on embedded processing of speech
Authors
xxv
and audio signals, he became in 2004 a research associate at the Department of Electrical Engineering in the University of Bremen. He finished his PhD thesis on automated systems for freight supervision at the end of 2009. His tasks inside the CRC 637 research cluster comprise of the development of the sensor and data processing system.
Jeken, Oliver Oliver Jeken studied industrial engineering and management at the University of Bremen. After his studies, he joined the newly founded work group of Prof. Dr.-Ing. Katja Windt for Global Production Logistics at Jacobs University Bremen as a research associate and PhD student. His main research interest is operations management and autonomous control.
Kopfer, Heiko Dipl.-Wirt.-Ing. Heiko Kopfer, born in 1982, finished his diploma in 2009 at the University of Siegen. Since 2009, he is a research assistant at the Institute of Machine Elements, Fastening Systems and Product Innovation (MVP).
Kopfer, Herbert Prof. Dr.-Ing. Herbert Kopfer has been managing the Chair of Logistics of the University of Bremen since its foundation in 1992. He is visiting professor at the University of Rennes. Currently, he is member of the Collaborative Research Center 637 “Autonomous Cooperating Logistics Processes – A Paradigm Shift and its Limitations” and principal investigator of two subprojects.
Korsmeier, Benjamin Dipl.-Kfm. Benjamin Korsmeier studied Economics at the University of Bielefeld from 2002 to 2008. During his studies, he emphasized on Marketing, Strategic Management, and Human Resources. In addition, he gained practical experiences by working for a medium-sized food-selling company for about 10 years. In November 2008, he started as a research associate in the department of Prof. Hülsmann in Bremen. His main field of interest is related to Strategic Management of Logistics Service Providers, especially in the field of “Strategic Positioning” and “Differentiation” by technologies.
xxvi
Authors
Kramer, Christoph Dipl.-Inf. Christoph Kramer studied Computer Science at the University of Bremen. He joined the Bremen Institute of Industrial Technology and Applied Work Science (BIBA) as a research scientist in 2009. Focus of the work of Mr. Kramer is the development of integration approaches for heterogeneous data sources. Mr. Kramer’s research area covers the definition of semantic and syntactic descriptions of heterogeneous data sources as well as implementations of transformation mechanisms in an ontology-based mediator system.
Kreowski, Hans-Jörg Hans-Jörg Kreowski is a professor of Theoretical Computer Science at the University of Bremen since 1982. He received his PhD and his habilitation from the Technical University of Berlin where he was a researcher and assistant professor from 1974 to 1982. He spent a sabbatical at IBM Research Center Yorktown Heights and was the first chair of the IFIP WG 1.3 (Foundations of Systems Specification). He is currently a team leader in the Collaborative Research Centre 637 on Autonomous Cooperating Logistic Processes. His main research areas are algebraic specification, graph transformation, picture generation, theory of concurrency, and formal methods in software engineering and logistics.
Kuske, Sabine Sabine Kuske is a lecturer at the Department of Computer Science in the University of Bremen in the north of Germany. She is a member of the Theoretical Computer Science Group headed by Prof. Dr. Hans-Jörg Kreowski. She received a PhD in Computer Science in 2000 on a thesis entitled "Transformation units – a structuring principle for graph transformation systems." Her research interests include all aspects of graph transformation with an emphasis on using graph transformation for visual modeling and verification of communities of autonomous units.
Lang, Walter Walter Lang studied physics at Munich University and received his Diploma in 1982 on Raman spectroscopy of crystals with low symmetry. His PhD in engineering at Munich Technical University was on flame-induced vibrations. In 1995, he became the head of the Sensors Department at the Institute of Micromachining and Information Technology of the Hahn-Schickard Gesellschaft (HSG-IMIT) in
Authors
xxvii
Villingen-Schwenningen, Germany. In February 2003, he joined the University of Bremen. He is the head of the Microsystems Center Bremen (MCB).
Langer, Hagen Dr. Hagen Langer is a senior researcher at the Artificial Intelligence Research Group of the University of Bremen. Jointly with the aforementioned authors, he leads the subproject Knowledge Management within the CRC 637. Besides knowledge representation and reasoning, an additional focus of his research is on natural language processing.
Luderer, Melanie A former member of the International Graduate School for Dynamics in Logistics at the University of Bremen and holder of a doctoral scholarship, Melanie Luderer is now a research assistant in the Theoretical Computer Science Group, also at the University of Bremen, which is headed by Prof. Dr. Hans-Jörg Kreowski. Her research interests lie in the verification of graph transformation processes, specifically their termination properties, and also the semantics of synchronous processes.
Malaka, Rainer Dr. Rainer Malaka is a professor and chair for Digital Media at the Department of Mathematics and Computer Science in the University of Bremen. He directs the TZI – Center for Computing and Communication Technologies concentrating on mobile solutions, digital reality, and adaptive communication. His research group works on projects dealing with mobile assistance systems, language understanding, geographical information systems, and computer vision.
Mironchenko, Andrii Andrii Mironchenko received the MSc degree in applied mathematics from Mechnikov Odessa National University in 2008. Since 2009, he is a research assistant at the Center for Industrial Mathematics in the University of Bremen. His research interests include stability of control systems, partial differential equations, and their applications to logistics and biology.
xxviii
Authors
Morales Kluge, Ernesto Dipl.-Wi.-Ing. Ernesto Morales Kluge works as a research scientist at the BIBA – Bremer Institut für Produktion und Logistik GmbH in the University of Bremen in the division Intelligent Production and Logistic Systems.
Naujok, Lars Lars Naujok studied Mathematics at the University of Bremen and received his diploma in 2008. He is a research assistant at the Center for Industrial Mathematics in the University of Bremen since 2009. His research interests include stability and control of dynamical systems, Lyapunov methods, observers, fault detection, and their applications to logistics.
Palafox-Albarran, Javier Javier Palafox-Albarrán received the BSc degree in Electronic Systems from the “Toluca Institute of Technology,” Mexico (ITT), in 2000. He finished his MSc in Information and Automation in 2009 at the University of Bremen. Since October 2010, he is a PhD student at the International Graduate School for Dynamics in Logistics in Bremen where he is doing research on the “Analysis and prediction of sensor and quality data in food transport supervision.”
Piotrowski, Jakub Dipl.-Inf. Jakub Piotrowski finished his study in Computer Science in 2005 and started as research scientist at the Bremer Institut für Produktion und Logistik GmbH (BIBA) in the University of Bremen, Germany. His main research activities focus on investigation of autonomous logistics processes in automobile logistics. Since 2009, Mr. Piotrowski is the managing director of the Collaborative Research Centre 637 "Autonomous Cooperating Logistic Processes – A Paradigm Shift and its Limitations" at the University of Bremen, Germany.
Porzel, Robert Dr. Robert Porzel is a senior researcher at the Digital Media Research Group in the University of Bremen. He is a researcher in the subproject Knowledge Management within the CRC 637. His research encompasses knowledge representation, contextual computing, and natural language processing.
Authors
xxix
Rekersbrink, Henning Dipl.-Ing. Henning Rekersbrink is a manager of the department “Dynamics and Complexity” at the BIBA – Bremer Institut für Produktion und Logistik GmbH in the University of Bremen in the division “Intelligent Production and Logistic Systems.”
Rippel, Daniel Dipl.-Inf. Daniel Rippel works as a research scientist at the BIBA – Bremer Institut für Produktion und Logistik GmbH in the University of Bremen in the division “Intelligent Production and Logistic Systems.” His research interests are in the field of autonomous control, model transformation, and collaborative decision management in nonhierarchical supply chains.
Ruthenbeck, Carmen Dipl.-Wi-Ing. Carmen Ruthenbeck works as a research scientist at the BIBA – Bremer Institut für Produktion und Logistik GmbH in the University of Bremen. She directs the department “Planning and Control Methods” and was involved in the transfer project “Sensor Systems for Storage Management.”
Scholz-Reiter, Bernd Prof. Dr.-Ing. Bernd Scholz-Reiter is the head of the department “Planning and Control of Production Systems” at the University of Bremen and managing director of the BIBA – Bremer Institut für Produktion und Logistik GmbH at the University of Bremen. He is the editor of the professional journals “Industrie Management and Productivity Management.” Prof. Scholz-Reiter is a corresponding member of C.I.R.P., the International Institution for Production Engineering Research and member of the Scientific Advisory Board of Bundesvereinigung Logistik e.V. He is also vice president of the German Research Foundation (DFG).
Schönberger, Jörn Dr. Jörn Schönberger was born in 1973 and graduated in Mathematics at Bremen University in 2000. Since then, he has been working at the Chair of Logistics receiving his PhD (Dr.rer.pol) in 2004 and finalizing his habilitation project in 2010. Dr. Schönberger’s current research is based on the project “Autonomous Adaptation of Vehicle Schedules” at the Collaborative Research Center 637.
xxx
Authors
Schuldt, Arne Dr. Arne Schuldt joined the Artificial Intelligence research group of Prof. Otthein Herzog at the Universität Bremen in 2006. In 2010, he received his PhD from the Universität Bremen with summa cum laude. His PhD thesis has been awarded with the Science Award for Logistics of the German Logistics Association (BVL) in 2010.
Schwientek, Anne Dipl.-Wi.-Ing. Anne Schwientek studied Industrial Engineering at TU Berlin and Karlsruhe Institute of Technology and received her diploma in 2009. Today, she is a PhD Student in the Systems Management working group of Prof. Dr. M. Hülsmann at Jacobs University Bremen. Her research interests include Green Logistics and Innovation Management.
Sowade, Steffen Dipl.-Ing. Steffen Sowade works as a research scientist at the BIBA – Bremer Institut für Produktion und Logistik GmbH in the University of Bremen in the division “Intelligent Production and Logistic Systems.” His research interests are in the field of autonomous control, service-oriented architectures, and protocols for manufacturing planning and control systems.
Teucke, Michael Dipl.-Wirtsch.-Ing. Michael Teucke works as research scientist at the BIBA – Bremer Institut für Produktion und Logistik GmbH in the University of Bremen. He was involved in the transfer project “Order Disposition in an Apparel Supply Chain.”
Thoben, Klaus-Dieter After finishing his studies in Mechanical Engineering, Klaus-Dieter Thoben worked as a research assistant at the Department of Production Engineering in the University of Bremen, where he received his Doctor of Engineering degree in 1989. He received the state doctorate (Habilitation) and the related “venia legendi” for the
Authors
xxxi
domain of Production Systems in 2002. In the same year, he was appointed to the professorship “IT Applications in Production Engineering” at the University of Bremen. Since 2003, he is director of BIBA (Bremen Institute for Production and Logistics GmbH) Department: ICT Applications in Production. In 2008, he was appointed to the professorship “Integrated Product Development” at the University of Bremen.
Totth, Caroline, von Caroline von Totth is a research assistant and doctoral candidate in the Theoretical Computer Science Group headed by Prof. Dr. Hans-Jörg Kreowski at the University of Bremen. Chief among her research interests is the rule-based generation of three-dimensional scenes followed closely by the formal modeling, simulation, and analysis of autonomous logistic processes by graph transformation as part of her work in the Collaborative Research Centre 637 on Autonomous Cooperating Logistic Processes.
Veigt, Marius Dipl.-Wi.-Ing. Marius Veigt works as a research scientist at the BIBA – Bremer Institut für Produktion und Logistik GmbH in the University of Bremen in the division “Intelligent Production and Logistic Systems.”
Wang, Xin Mr. Xin Wang is a PhD student at the Chair of Logistics in the University of Bremen. He holds a Master degree of Industrial Engineering and Management from the University of Bremen. His main research and work interests include development and implementation of heuristic algorithms applied to transportation planning and scheduling as well as modeling and simulation of horizontal collaboration among freight carriers.
Warden, Tobias Tobias Warden received his diploma degree in Computer Science from the University of Bremen in 2007 with a thesis on spatio-temporal analysis of dynamic scenes in the RoboCup domain. He joined the artificial intelligence group at the
xxxii
Authors
University of Bremen as a research assistant in 2008 and is now affiliated to the CRC 637. Current research interests span distributed knowledge management and collaborative multi-agent learning.
Wenning, Bernd-Ludwig Dr. Bernd-Ludwig Wenning is a postdoctoral researcher at the Communication Networks working group (ComNets), University of Bremen, where he received his Dr.-Ing. degree in 2009. As a member of the CRC637 “Autonomous Cooperating Logistic Processes,” his research focus is on context-based routing in logistics networks and in communication networks.
Windt, Katja Katja Windt is a professor of Global Production Logistics at Jacobs University Bremen. She is heading a partial project of the Collaborative Research Center 637 “Autonomous Logistics.” The aim is to investigate and develop methods for an autonomous product construction cycle enabling logistic objects to decide between its next production step alternatives. Further research interests of Katja Windt are development of production planning and control methods, design of distribution networks, graph coloring dynamics, auditory display in logistics, and biological inspired production control.
Zabel, Christian Christian Zabel has studied Computer Science at the University of Bremen and is currently working as a research scientist at the Bremer Institut für Produktion und Logistik GmbH. His research interests include the autonomous routing of logistics objects.
Chapter 1
Contributions and Limitations of Autonomous Cooperation and Control in Logistics Michael Hülsmann, Bernd Scholz-Reiter, and Katja Windt
1.1 In Search of the Optimal Degree of Autonomous Cooperation and Control in Logistics The realization of autonomous cooperation and control in logistics is not a question of either external control or autonomous control. Rather, it implies the intention to increase the degree of autonomous cooperation and control [9]. This means that at least one of the constitutive characteristics – autonomy, decentralized decisionmaking, interaction, heterarchy and non-determinism [24] – is ceteris paribus intensified. Technologies (e.g. the intelligent container [15, 16]) or methodologies (e.g. collaborative transportation planning [17]), whose implementation or usage does augment these characteristics in logistics processes can therefore be regarded as enablers for autonomous cooperation and control. In contrast, organizational approaches, technological solutions or planning and controlling routines, which lead to a diminution of the constitutive characteristics, can be appraised as impediments for autonomous cooperation and control in logistics. Consequently, two main questions arise: Firstly, which degree of autonomous cooperation and control is feasible? Secondly, which degree of autonomous cooperation and control is reasonable? In other words: which technologies and methodologies are able to increase the degree of autonomous cooperation, how effective are these measures and what is the adequate implementation level in logistics systems? The underlying purpose is to find the “optimal degree” of autonomous cooperation and control of logistics processes. This leads to the necessity to know the M. Hülsmann (B) SyM, Jacobs University Bremen, Campus Ring 1, 28759 Bremen e-mail:
[email protected] B. Scholz-Reiter IPS, BIBA, Universität Bremen, Hochschulring 20, 28359 Bremen e-mail:
[email protected] K. Windt GPL, Jacobs University Bremen, Campus Ring 1, 28759 Bremen e-mail:
[email protected] M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_1,
1
2
M. Hülsmann et al.
drivers for the optimal degree and therefore to know the enablers and barriers for autonomous cooperation and control in logistics systems – on the level of organization and management, information and communication as well as material flow [22]. Hitherto conducted research shows first promising findings: On the organization and management level, for example, autonomous cooperation and control can contribute to the flexibility [10, 25] and the robustness [12] of a company’s structures and processes and to a transportation system’s efficiency [2, 7]. Furthermore, share prices of logistics companies indicates that the implementation of autonomous cooperation-enabling technologies has a positive influence on the company value [4]. On the information and communication level, increasing the degree of autonomous cooperation and control of logistics processes can decrease communication costs and energy consumption [14]. Also, the quality of transported food, for example, can be improved by observing the conditions of the transport process and adjusting the transport routes [6]. On the material flow level, implementing autonomous cooperation and control in logistics can improve a production system’s performance by increasing the ability to cope with complexity [21] and can lead to a higher achievement of logistic objectives (e.g. lead times, due-date reliability) [3]. However, recent research shows also barriers and limitations, which prevents the realization of a higher degree of autonomous cooperation and control of logistics processes: On the organization and management level, research indicates that implementing autonomous cooperation-enabling technologies bears higher risks than the implementation of other technologies [4], for example strategic risks, legal risks or operational risks [11]. On the information and communication level, technical limitations can be observed, such as the problem of radio propagation, which describes an attenuation of radio waves during the transmission from the sender to the receiver [1]. On the material flow level limitations result for instance from the risk that autonomous control methods might cause sudden and unexpected changes of the respective logistics system’s behavior. Under certain conditions an intrinsic dynamic behavior and even chaos can occur [20]. In consequence, both questions – to which degree it is possible and to which degree it makes sense to increase the degree of autonomous cooperation and control of logistics processes – have been addressed in former research. However, it is not clear whether the contributions or the limitations overweigh. Additionally, these results cannot be generalized because they focus on individual problems based on individual applications and theoretical foundations [25]. Hence, there is still a gap in research regarding the general contributions and limitations of autonomous cooperation and control of logistics processes. This would substantially contribute to the answer of the central ideal question regarding a general, optimal and feasible degree of autonomous cooperation and control in logistics. Correspondingly, the underlying aim is to improve the capability to cope better with complexity and dynamics of logistics systems. Therefore, it is essential to gain insights into the benefits und feasibility of autonomous cooperation and control in logistics. This edited volume approaches this question from four different
1
Contributions and Limitations of Autonomous Cooperation and Control in Logistics
3
perspectives on contributions and limitations: organizational, methodical, technical, and practical. The organizational perspective comprises four views: institutional, functional, instrumental, and process-related [8]. Each of these views concentrates on different aspects of organization such as the system itself, the system’s structure, measures or processes of organizational design. Therefore, scientific research investigating organizational contributions and limitations of autonomous cooperation and control in logistics needs to consider these views in order to be able to discuss the effects on the system’s goal achievement. The methodical perspective implies in a broader view mainly three aspects: selecting agents, defining their goals [18], and modeling a strategy. In a closer view, Windt et al. (2010) define autonomous control methods as “generic algorithms that describe how logistics objects render and execute decisions by their own” [23]. For the purpose to investigate contributions and limitations of autonomous cooperation and control in logistics, the broader view should be considered in order to include all of influencing aspects. The technical perspective mainly focuses on developing and improving features of information and communication technologies such as RFID, GPS, sensor networks, etc., and therefore, on physically realizing autonomous control methods. This includes issues such as data integration or multi agent systems. Therefore, the technical perspective provides physical results, which enable logistics objects to decide decentralized. The practical perspective, finally, refers to the actual application of autonomous cooperation and control methods and technologies in logistics practice. Implementing autonomous cooperation and control into industrial routines might produce non-intended results. Therefore, the validity of theoretical scientific findings needs to be tested in the “real-world” in order to be able to give recommendations for logistics management concerning the implementation of autonomous cooperation and control. All in all, the major research question of this edited volume is the following: What are the major driving forces – fostering as well as hindering determinants – for the feasible and reasonable utilization of the idea of autonomy in the decentralized decision-making of logistics objects in complex and dynamic supply networks – on the level of organization and management, information and communication and material flow?
1.2 Aims In order to provide a comprehensive and profound knowledgebase for the raised question, the edited volume aims for the identification, description, analysis, and evaluation of contributions & limitations of autonomous cooperation and control from organizational, methodical, technical, and practical perspectives. Therefore, it is intended to develop fundamental contributions to the achievement of the
4
M. Hülsmann et al.
following research aims, which are connected with the overall objective of the research related to autonomous cooperation and control in logistics. From an organizational perspective: (a) Describing possible organizational determinants, which enable the implementation of the concept of autonomous cooperation and control in logistics. (b) Analyzing causal relationships between implementing autonomous cooperation and control and organizational structures, processes, activities, and tasks in logistics companies and networks in both directions. (c) Deducing recommendations concerning the design of organizational structures and processes in order to implement autonomous cooperation and control technologies. From a methodical perspective: (a) Developing a general setting of methodologies for modeling autonomous cooperation and control in logistics, which is compatible to existing approaches. (b) Finding a systematic approach for deciding, which autonomous cooperation and control method is best suitable for a certain logistics scenario. (c) Determining analytically the boundaries of computational capabilities depending on the autonomous cooperation and control method used. From a technical perspective: (a) Depicting today’s technical possibilities to apply the concept of autonomous cooperation and control on logistics processes. (b) Investigating technical pre-conditions, which need to be fulfilled in order to enable the implementation of autonomous cooperation and control in logistics. (c) Examining systematically technical restrictions, which limit the possible and feasible degree of autonomous cooperation and control in logistics. From a practical perspective: (a) Documenting experiences from first applications of autonomous cooperation and control in example production and distribution logistics scenarios. (b) Deriving from these experience, on the one hand, the potential of autonomous cooperation and control to improve logistics performance. (c) Investigating, on the other hand, impediments of autonomous cooperation and control technologies and methods after implementation into daily industrial routines.
1.3 Structure and Results In order to contribute to the research aims addressed above, the edited volume is divided into four corresponding sections: Organizational contributions and limitations Methodical contributions and limitations
1
Contributions and Limitations of Autonomous Cooperation and Control in Logistics
5
Technological contributions and limitations Practical contributions and limitations
The second chapter “Organizational Contributions and Limitations” investigates the impact of autonomous cooperation and control on the organization of logistics systems. Considering four different perspectives of organization (institutional, functional, instrumental, and process-related [8]), this chapter analyzes how the design of structures and processes is affected by an increased autonomy, interaction, and decentralized decision-making in order to determine effects on the achievement of logistics companies’ and systems’ goals. This is a fundamental perspective on autonomous cooperation and control since this concept is rooted in the principle of self-organization [13], which analyzes how ordered structures emerge in complex and dynamic systems [5, 19]. The chapter’s findings show, for example, that logistics companies can use autonomous cooperation-based technologies in order to differentiate from competitors by creating additional customer value. Therefore, autonomous cooperation and control can have a positive impact on the strategic positioning of logistics companies. Furthermore, the effects of different design options of transport chain organization and transportation planning on efficiency and expected revenues are presented. In addition, this chapter shows that autonomous control positively affects a logistics system’s (exemplarily a production network) effectiveness, if information about the goods flows is not available. This allows to investigate, how an increase of the degree of autonomous cooperation and control affects the organization of logistics systems. The third chapter “Methodical Contributions and Limitations” discusses aspects of autonomous control methods in logistics. This implies a modeling methodology and preconditions and implications of different methods, but also computability in order to assess the causal relationships between different logistics parameters, autonomous control methods and the resulting performance. The methodical perspective is an essential component for applying the concept of autonomous cooperation and control in logistics [22], since methods describe, for example, how autonomous logistics objects decide and act. The findings for instance, analyze the performance of different methods depending on dynamics and complexity of the logistics system. Generally, autonomous control methods are superior to conventional methods, if dynamics and complexity increases. However, it is shown that multiple methods within one system can reduce the performance although single methods perform well. This contributes, on the one hand, to the aim to systematically map suitable methods depending, for example, on scenario parameters, and, on the other hand, to the aim to determine the optimal degree of autonomous cooperation and control. In addition, a framework is developed for computational modeling in order to handle the computational complexity of autonomous control methods. This contributes to the aim to develop a general setting of methodologies. The fourth chapter “Technological Contributions and Limitations” considers technological requirements, which are necessary to realize autonomous control methods in logistics. The technological perspective is another essential component
6
M. Hülsmann et al.
for applying the concept of autonomous cooperation and control in logistics since the use of information and communication technologies enables logistics objects to interact, to obtain necessary information, and to decide autonomously [24]. Findings are presented, for example, regarding reliability, range of communication, and the required energy resource of semi-passive RFID tags in order to realize an “intelligent” container. Furthermore, effects of different data integration approaches on the performance (e.g. timeliness, robustness) are discussed, which limit the possible and feasible degree of autonomous cooperation and control. These results contribute to the three beforehand mentioned aims as, first, technical possibilities are considered, second, technical requirements are analyzed, and third, technical limitations are investigated by comparing possibilities and requirements. The fifth chapter “Practical Contributions and Limitations” investigates the realizability of autonomous cooperation and control-based technologies and methods in logistics practice in order to consider aspects that appear if the respective technologies and methods are implemented into daily industrial routines and if “real” data are used. Since autonomous cooperation and control is a relatively new concept, applications in the field are rare. Therefore, there is a need to investigate the usability and feasibility of implementing the concept of autonomous cooperation and control into logistics systems and processes. The results show that the concept of autonomous cooperation and control is, basically, applicable to logistics and can improve logistics performance. Furthermore, the assumption is validated that autonomous control methods need to be customized for a certain logistics scenario and cannot be recommended generally. For example, the technological infrastructure has to fit the requirements of the chosen methods. Also, the findings of chapter five show the necessity that relevant information is made available. Therefore, this chapter documents implementation experiences and contributes to the aims to derive the potential but also boundaries of autonomous cooperation and control in logistics practice.
References 1. Becker M, Yuan S, Jedermann R, Timm-Giel A, Lang W, Görg C (2009) Challenges of applying wireless sensor networks in logistics. CEWIT 2009. Wireless and IT driving healthcare, energy and infrastructure transformation 2. Bloos M, Kopfer H (2009) Efficiency of transport collaboration mechanisms. Comm SIWN 6(1):23–28 3. Böse F, Piotrowski J, Scholz-Reiter B (2008) Autonomously controlled storage management in vehicle logistics – applications of RFID and mobile computing systems. Int J RF Tech ResAppl 1(1):57–76 4. Cordes P, Illigen C, Hülsmann M (2011) Effects of autonomous cooperation-enabling technologies on the growth of share prices of logistics enterprises. In: Hülsmann M (ed.) Research contributions to strategic management 27. Jacobs University Bremen, Bremen 5. Flaemig M (1998) Naturwissenschaftliche weltbilder in managementtheorien: chaostheorie, selbstorganisation, autopoiesis. Campus-Verl, Frankfurt am Main et al. 6. Hentschel D, Jedermann R, Lang W (2009) Autonomous cooperating processes for the improvement of food quality. In: Proceedings sensor 2009 volume 2. Sensor + Test Conference 2009AMA Service GmbH, Wunstorf. 447
1
Contributions and Limitations of Autonomous Cooperation and Control in Logistics
7
7. Hongler M, Gallay O, Hülsmann M, Cordes P, Colmorn R (2010) Centralized versus decentralized control - a solvable stylized model in transportation. Phys A 389(19):4162–4171 8. Hülsmann M (2003) Management im orientierungsdilemma: unternehmen zwischen effizienz und nachhaltigkeit. Dt. Univ.-Verl 9. Hülsmann M, Grapp J (2006) Monitoring of autonomous cooperating logistic processes in international supply networks. In: Pawar KS et al. (eds.) Conference proceedings of 11th international symposium on logistics (11th ISL), pp 113–120. Loughborough, United Kingdom 10. Hülsmann M, Grapp J, Li Y (2006) Strategic flexibility in global supply chains – competitive advantage by autonomous cooperation. In: Pawar KS et al. (eds.) Conference proceedings of 11th international symposium on logistics (11th ISL), pp 494–502. Loughborough, United Kingdom 11. Hülsmann M, Illigen C, Korsmeier B, Cordes P (2010) Risks resulting from autonomous cooperation technologies in logistics. In: Li Y, Desheng W (eds.) Proceedings of the 2010 IEEE international conference on advanced management science, pp 422–426 12. Hülsmann M, Scholz-Reiter B, deBeer C, Austerschulte L (2008) Effects of autonomous cooperation on the robustness of international supply networks – contributions and limitations for the management of external dynamics in complex systems. In: Haasis H, Kreowski HJ, ScholzReiter B (eds.) Dynamics in logistics – proceedings of the 1st international conference on dynamics in logistics, August 28th – 30th, pp 241–250, 2007, Bremen, Germany 13. Hülsmann M, Wycisk C, Agarwal R, Grapp J (2007) Prologue to autonomous cooperation – the idea of self-organisation as its basic concepts. In: Hülsmann M, Windt K (eds.) Understanding autonomous cooperation & control – the impact of autonomy on management, information, communication, and material flow, pp 23–44. Springer, Berlin 14. Jaffari AM, Sklorz A, Lang W (2009) Energy consumption comparison between autonomous and central wireless sensor networks. Comm SIWN 6(6):166–170 15. Jedermann R, Lang W (2008) The benefits of embedded intelligence – tasks and applications for ubiquitous computing in logistics. In: Floerkemeier C, Langheinrich M, Fleisch E, Mattern F, Sarma SE (eds.) The internet of things – first international conference, IOT 2008, Zurich, Switzerland, March 26-28, 2008. Proceedings, pp 105–122. Springer, Berlin, Heidelberg, New York 16. Jedermann R, Ruiz-Garcia L, Lang W (2009) Spatial temperature profiling by semi-passive RFID loggers for perishable food transportation. Comput Electron Agr 65(2):145–154 17. Krajewska M, Kopfer H (2006) Collaborating freight forwarding enterprises - request allocation and profit sharing. OR Spectrum 3(28):301–317 18. Nyhuis P, Wiendahl H, Rossi R (2009) Fundamentals of production logistics: theory, tools and applications. Springer, Berlin 19. Paslack R (1991) Urgeschichte der Selbstorganisation: zur Archäologie eines wissenschaftlichen Paradigmas. Vieweg, Braunschweig 20. Scholz-Reiter B, Görges M, Jagalski T, Mehrsai A (2009) Modelling and analysis of autonomously controlled production networks. In: Proceedings of the 13th IFAC symposium on information control problems in manufacturing (INCOM 09) Moscow, Russia, 850 21. Scholz-Reiter B, Görges M, Philipp T (2009) Autonomously controlled production systems – Influence of autonomous control level on logistic performance. CIRP Ann Manuf Tech 58(1):395–398 22. Scholz-Reiter B, Windt K, Freitag M (2004) Autonomous logistic processes: new demands and first approaches. In: Monostori L (ed.) Proceedings of the 37th CIRP international seminar on manufacturing systems, Budapest, 357–362 23. Windt K, Becker T, Jeken O, Gelessus A (2010) A classification pattern for autonomous control methods in logistics. Logist Res 2(2):109–120 24. Windt K, Hülsmann M (2007) Changing paradigms in logistics – understanding the shift from conventional control to autonomous cooperation and control. In: Hülsmann M, Windt K (eds.) Understanding autonomous cooperation & control – the impact of autonomy on management, information, communication, and material flow, pp 1–16. Springer, Berlin 25. Wycisk C (2009) Flexibilität durch selbststeuerung in logistischen systemen: entwicklung eines realoptionsbasierten bewertungsmodells. Gabler
Part I
Organizational Contributions and Limitations
Chapter 2
Approaches to Organizational Contributions and Limitations of Autonomous Cooperation and Control in Logistics Michael Hülsmann, Anne Schwientek, and Philip Cordes
The concept of autonomous cooperation and control refers back to the idea of self-organization, which describes, how system structures and processes emerge without an external influence [1–3]. A logistics system – a logistics company or an entire supply network – is autonomously cooperating and controlled when its logistics objects (e.g. containers, packets) interact with each other and are enabled to decide autonomously and decentralized from a central control entity (e.g. on transport routes or means of transport) [4–6]. The organization of the underlying logistics systems can be regarded from an institutional, a functional, an instrumental and a process-related perspective [7]. Each of them is affected when the degree of autonomous cooperation and control of logistics processes is changed. The institutional perspective understands organization as a social system with a specific purpose [1, 8–11]. It allocates the processes, the tasks and instruments to persons or logistics objects (based on [7]). Companies or entire supply networks – as social systems in terms of business enterprises or networks of cooperating enterprises – do generally follow the purpose of long-term profit maximization [12]. A decentralization of decision-making shifts the allocation of certain tasks from a central management to the single objects [13]. The objects themselves conduct the allocation of processes as well the selection of corresponding instruments [5]. Hence, the institutional setting of the logistics system’s organization is directly affected. The functional perspective describes organization as goal-oriented control of a system’s actions [8, 10, 14]. It represents the content of tasks within logistics processes [7]. In an autonomously controlled logistics system, not a central management determines the content of the logistics objects’ tasks but the objects themselves. They interact with each other and decide upon their next steps autonomously [13]. Consequently, an alteration of the degree of autonomous cooperation and control influences also the functional perspective of a company’s or network’s organization. M. Hülsmann (B), A. Schwientek, and P. Cordes Systems Management, International Logistics, School of Engineering and Science, Jacobs University Bremen, Bremen, Germany e-mail:
[email protected] M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_2,
11
12
M. Hülsmann et al.
The instrumental view understands organization as measures or regulations for designing a system’s structures (e.g. [15, 16]). It describes the ‘tools’ used for the design of the organization [7]. On the one hand, autonomous cooperation and control can be utilized as tool itself, since it enables an autonomous emergence of systems’ structures. On the other hand, a higher degree of autonomous cooperation shifts the selection of tools from a central management to the single logistics objects [5]. Thereby, the instrumental setting of a company’s or a supply network’s organization is also changed by an increase of autonomous cooperation and control. Finally, from a process-related perspective organization is understood as the sequence of activities for designing regulations that enable a goal-oriented cooperation of a system’s participants (e.g. [17, 18]). In other words, it describes the structuring of sequences of tasks that are necessary in order to achieve the system’s goals [7] – in logistics, for instance, due-date reliability [19]. In autonomously cooperating logistics systems, the allocation of these sequences results from the interaction and autonomous decision-making of the logistics objects [20]. Compared to an externally controlled logistics system, the processes might be allocated differently. Hence, the process-related perspective of the logistics system’s organization is directly affected. Considering these first and exemplary thoughts on institutional, functional, instrumental and process-related effects of autonomous cooperation and control, it can be assumed that an implementation of associated technologies (e.g. the intelligent container [21]) or methodologies (e.g. groupage systems [22]) affects the organization of logistics companies and entire supply networks in several ways. Therefore, the question arises, in what way autonomous cooperation and control contributes to or limits the organization of logistics systems to follow their purposes and achieve their settings of goals? Correspondingly, the central aim of this chapter is to reveal the effects of increasing the degree of autonomous cooperation and control of logistics processes on the organization of the underlying logistics systems. In order to illustrate the range of impacts on the organization, different perspectives of analysis should be applied. Therefore, institutional, instrumental, functional and process-related understandings of organization will be taken as the basis for the analysis. Analytically, the aim is to reveal causal relationships between the implementation of autonomous cooperation and control-based technologies and the organization of logistics systems. In consequence, practical implications for the goal-oriented organizational design of logistics companies and entire supply networks shall be deduced. Furthermore, it is intended to reveal the spectrum of alternative applications of autonomous cooperation and control. Thereby, practitioners shall be supported in getting used to the concept and to get to know its contributions and limitations in logistical settings. Michael Hülsmann, Anne Schwientek, Benjamin Korsmeier, and Linda Austerschulte discuss in their contribution “Creating Customer Value in Logistics – Contributions and Limitations of Autonomous Cooperation-based Technologies” the implications of the intelligent container as an example of autonomous cooperationbased technology for creating a unique value for the customers of a logistics company. They analyze the implications of offering services based on the intelligent
2
Approaches to Organizational Contributions and Limitations
13
container’s features on customer value dimensions. Thereby, they mainly take a functional perspective as they consider the goal-oriented decision of implementing the intelligent container in order to create additional customer value. Sergey Dashkovskiy, Andrii Mironchenko, and Lars Naujok examine in their contribution “Autonomous and Central Control of Production Networks” the implications of changing the degree of autonomous control on the stability of production networks. By implementing central planning methods into mathematical models (based on optimal control theory) they find answers regarding the effective control of production networks as well as the contributions and limitations of central and autonomous control methods. Therefore, they focus on an institutional and an instrumental perspective as they consider a production network as a system with certain properties (Dinstitutional) and the implications of different control methods (as tools for designing the system structures) on the system goal stability (Dinstrumental). In their contribution “Approaching the application borders of network capacity control in road haulage”, Jörn Schönberger and Herbert Kopfer investigate effects of decentralized transport chain organization. By using capacity control mechanisms, actors in transportation (as freight forwarders and carriers) decide autonomously on the acceptance of transport requests and thus on the consumption of transport resources. In this context, Schönberger and Kopfer analyze contributions and limitations to the expected revenues of basic models and of capacity control strategies depending on demand forecast accuracy. The perspective chosen here/in this paper is a functional view, since effects of different organizational models on expected revenues are discussed. However, they also consider the instrumental perspective since the use of capacity control can be regarded as a measure to design the organizational structure. Heiko Wieland Kopfer, Herbert Kopfer, and Xin Wang analyze in their contribution “Limit and Degree of Autonomy in Groupage Systems – Challenges, Chances and Barriers for Horizontal Cooperation in Operational Transportation Planning” groupage systems of small and medium-sized transportation companies. These companies can improve their competitiveness by cooperating horizontally (e.g. exchanging transportation requests, collaborative vehicle routing). In this context, Kopfer, Kopfer, and Wang investigate the efficiency of groupage systems depending on the intensity of cooperation and each partner’s degree of autonomy. The instrumental perspective serves as a basis since they investigate the impact of different organizational design options (as organizational measures) on the degrees of intensity and autonomy within the group of cooperating companies. Arne Schuldt, Jan Ole Berndt, and Otthein Herzog also investigate different organizational structures in logistics while focusing on a transport scenario of sea containers. In their contribution “The Interaction Effort in Autonomous Logistics Processes – Potential and Limitations for Cooperation”, they discuss the effects of autonomous control and different organizational structures on the required computational and interaction effort for decentralized decision-making. While an increase of autonomous control decreases the computational effort, it increases the necessary interaction. Thus a functional perspective is chosen as they analyze the impact
14
M. Hülsmann et al.
of organizational structures on certain goals. Furthermore, they also consider an instrumental perspective as they discuss different methods of cooperation such as team formation.
References 1. Bea FX, Göbel E (2006) Organisation: theorie und gestaltung. 3., neu bearb. Lucius & Lucius, Aufl ed. Stuttgart 2. Heylighen F (2003) The science of self-organization and adaptivity. In: Kiel LD (ed) The encyclopedia of life support systems. EOLSS Publishers, Oxford 3. Probst GJB (1987) Selbst-organisation: ordnungsprozesse in sozialen systemen aus ganzheitlicher sicht. Parey, Berlin 4. Hülsmann M, Windt K (eds) (2007) Understanding autonomous cooperation & control – the impact of autonomy on management, information, communication, and material flow, 1st edn. Springer, Berlin 5. Scholz-Reiter B, Windt K, Freitag M (2004) Autonomous logistic processes: new demands and first approaches. In: Monostori L (ed.) Proceedings of the 37th CIRP international seminar on manufacturing systems, Budapest. 357–362 6. Freitag M, Herzog O, Scholz-Reiter B (2004) Selbststeuerung logistischer Prozesse – ein Paradigmenwechsel und seine Grenzen. Ind Manag 20(1):23–7 7. Hülsmann M (2003) Management im orientierungsdilemma: unternehmen zwischen effizienz und nachhaltigkeit. Dt. Univ.-Verl, Wiesbaden 8. Laux H, Liermann F (2005) Grundlagen der Organisation: Die Steuerung von Entscheidungen als Grundproblem der Betriebswirtschaftslehre, 6th edn. Springer, Berlin [u.a.] 9. Jost P (2000) Organisation und koordination: eine ökonomische einführung. Gabler, Wiesbaden 10. Schertler W (1998) Unternehmensorganisation – lehrbuch der organisation und strategischen unternehmensführung, 7th edn. Oldenbourg, München [u.a.] 11. Milgrom P, Roberts J (1992) Economics, organization, and management. Prentice Hall, Englewood Cliffs 12. Wöhe G, Döring U (2000) Einführung in die allgemeine Betriebswirts, 20th edn. Franz Vahlen, München 13. Windt K, Hülsmann M (2007) Changing paradigms in logistics – understanding the shift from conventional control to autonomous cooperation and control. In: Hülsmann M, Windt K (eds.) Understanding autonomous cooperation & control – the impact of autonomy on management, information, communication, and material flow, pp 1–16. Springer, Berlin 14. Schulte-Zurhausen M (1999) Organisation, 2nd edn. Vahlen, München 15. Remer A (1997) Organisationslehre, 4th edn. REA-Verl. Managementforschung, Bayreuth 16. Hill W, Fehlbaum R, Ulrich P (1994) Organisationslehre: ziele, instrumente und bedingungen der organisation sozialer systeme, 5th edn. Haupt, Bern [u.a.] 17. Hicks HG, Gullett CR (1975) Organizations – theory and behaviour, 1st edn. McGraw-Hill, New York 18. Krüger W (1993) Organisation der unternehmung, 2nd edn. Kohlhammer, Stuttgart 19. Beamon BM (1999) Measuring supply chain performance. Int J Oper Prod Manag 19(3): 275–292 20. Windt K, Böse F, Philipp T (2008) Autonomy in production logistics – identification, characterisation and application. Int J Robot CIM 24(4):572–578 21. Gehrke JD, Behrens C, Jedermann R, Morales-Kluge E (2006) The intelligent container – toward autonomous logistic processes KI 2006 Demo Presentations, 15–18 22. Kopfer H, Schönberger J, Bloos M (2010) Groupage systems – collaborative request fulfillment in road haulage: A procedural view. In: Schönberger R, Elbert R (eds.) Dimensionen der logistik: funktionen, institutionen und handlungsebenen, pp 787–804. Gabler, Wiesbaden
Chapter 3
Creating Customer Value in Logistics: Contributions and Limitations of Autonomous Cooperation-Based Technologies Michael Hülsmann, Anne Schwientek, Benjamin Korsmeier, and Linda Austerschulte
3.1 Introduction Logistics service providers face various challenges like changes in customer expectations or technological breakthroughs [1]. One major challenge results from the homogeneity of many products logistics service providers offer [2]. Additionally, customers demand fast, reliable, customized, and cost-effective logistics processes more and more [3]. The possibility to fulfill suchlike customer demands requires a high flexibility of logistics service providers to adapt their services [4]. In order to establish and maintain a unique selling proposition logistics service providers have to separate themselves from their competitors by offering a different or additional service [5]. This implies that logistics service providers have to make themselves distinguishable from their competitors, which can be achieved by creating a unique value to their customers. Customer value can be viewed at as being composed of two components: benefits and sacrifices [6]. Benefits – as utility value – increase customer value. In contrast, sacrifices – as the price the customer has to pay – decrease customer value [7]. Therefore, it is necessary to analyze benefits and sacrifices in order to determine their net effect and to minimize the risk of decreasing customer value. If customers think they obtain a greater value from a company than from other companies, they are more likely to authorize the company’s services [8]. In this context, Butz and Goldstein refer to creating customer value as establishing an emotional bond between customer and supplier, which leads the customer to buy repeatedly from this supplier and recommend it to others [9]. Thus, if a logistics service provider can create more value to its customers – which is perceived by the customers [10] – the logistics service provider increases its chances to receive orders and, eventually, to stay in business. Thus, if a logistics service provider should differentiate from its competitors by increasing additional customer value in order to M. Hülsmann (B), A. Schwientek, B. Korsmeier, and L. Austerschulte Systems Management, International Logistics, School of Engineering and Science, Jacobs University Bremen, Bremen, Germany e-mail:
[email protected] M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_3,
15
16
M. Hülsmann et al.
stay competitive, the question arises: How can logistics service providers increase customer value? One option is to offer new services. These services can be based on using innovative technologies since they can help a company to offer services, which are different from its competitors’ services [11] (as long as the competitors have no access to these technologies or at least don’t use them). Examples for new technologies are those based on the principle of autonomous cooperation. Autonomous cooperation “. . . describes processes of decentralized decision-making in heterarchical structures. It presumes interacting elements in non-deterministic systems, which possess the capability and possibility to render decisions.” [12]. An example for autonomous cooperation-based technologies is the intelligent container, which offers features like real-time information and autonomous reaction to critical events during transportation [13–15]. Thus, the question arises: Can autonomous cooperation-based technologies be used to create customer value? Therefore, this article aims to identify contributions and limitations of the intelligent container (as a representative example of autonomous cooperation-based technologies) for generating customer value. For this purpose, the concept of customer value in a logistics environment is described in terms of value dimensions (Sect. 3.2). Furthermore, the intelligent container as an autonomous cooperationbased technology is described with respect to its features (Sect. 3.3). Next, the (positive and negative) effects of the intelligent container’s features on the customer value dimensions are analyzed by considering a transport service conducted with the intelligent container as an offered product (Sect. 3.4). This analysis can be used as a basis for logistics service-providers’ evaluations about whether to use intelligent containers to gain strategic advantage over their competitors.
3.2 Customer Value in Logistics Business Models Markets for logistics services are characterized by new challenges and dynamics, triggered by a change in external conditions [3]. These changes result amongst others from increased customer demands [16] as well as from increased competition in logistics markets [1]. Therefore, these challenges have to be addressed through the range of offered services and incorporated in the marketing activities of logistics service providers in order to find an appropriate position in the market [17]. Moreover, an alteration in the offered services differentiates a company’s product mix as well as the company itself from competitors [5]. Therefore, the question arises: How can a company (especially a logistics service provider) differentiate from its competitors? One possibility for differentiating a company’s products or services from competitors in order to achieve competitive advantage is to create additional value for the customers [10, 18]. Customer value is established if customers perceive the services provided by an organization as more beneficial than the related costs (net customer value) in comparison to competitive offers [9, 19]. In order to meet changing customer demands, logistics service providers have to adapt their service range [4]
3
Creating Customer Value in Logistics: Contributions and Limitations
17
towards services that are perceived by customers as more beneficial than the related costs, i.e. new or improved services with a potential to create increased customer value. Moreover, the resulting net customer value has to be higher than the perceived value of competitors’ services in order to differentiate the firm from its competitors [9]. However, logistics service providers have to consider competitors, which will also strive to adapt their service program to customer needs. Thus, a logistics service provider needs to be capable of offering services, which not only satisfy customer demands but also maximize customer value. One approach to address this challenge can be firm-specific capabilities [20] (e.g. capability to provide real-time information), based on technologies (e.g. new information and communication technologies), which are either not used or used in another way by competitors [21]. These capabilities can lead to new or improved services (e.g. transportation processes with the availability of real-time information). In the field of logistics autonomous cooperation-based technologies, like the intelligent container, can be one approach to be used in this context. Therefore, it has to be revealed, on the one hand, how autonomous cooperationbased technologies contribute to the creation of customer value. On the other hand, it has to be investigated whether the use of suchlike technologies also limits the creation of customer value and consequently the options to differentiate from competitors. In order to estimate the net customer value of an offered service, companies have to be aware of the value dimensions, which are perceived as benefits or as sacrifices in the customer’s mind [7, 9, 22–24]. Thus, logistics service providers need to adapt their product or service portfolio with respect to cost effective dimensions. Since our research focuses on logistics services in which the considered product is a service, the terms ‘service’ and ‘product’ will be used synonymously in this contribution. Regarding the dimensions of customer value, Khalifa [7] developed a value exchange model based on former research by [22, 24–27]. Following Khalifa [7], net customer value consists of the difference between total customer value (benefits) and total customer costs (sacrifices) (see Fig. 3.1). Note that total customer value is composed of a utility value and a psychic value. While the utility value results from the core service, the psychic value is “above the pure utilitarian value” and serves for psychic need fulfillment (e.g. image, status, affinity) [27]. Concerning utility value in a logistics context, the service attributes that lead to the creation of customer utility can be described by the seven R’s: the ability of a logistics company to deliver the right amount of the right product at the right place at the right time in the right condition at the right price with the right information [28]. Thus, the fulfillment of these service attributes creates utility value for the company demanding the service (e.g. producer of fruits). Total customer costs can be divided into a financial costs component (cost for the service itself, i.e. the price paid) and a non-financial costs component (costs occurring before, at, or after the service use, e.g. search for the right logistics service provider) [7]. Thus, following Khalifa [7], customer value is composed of four dimensions (utility value, psychic value, financial costs and non-financial costs). These criteria
18
M. Hülsmann et al.
Psychic Value
Utility Value
Financial Costs
Total Customer Value
Non-Financial Costs
Total Customer Costs
Net Customer Value
Fig. 3.1 Dimensions of net customer value (Own creation according to [7])
may be used for analyzing the implications of new services based on autonomous cooperation-based technologies relevant to customer value. However, to be able to conduct such analyses, also autonomous cooperation-based technologies need to be described with respect to their features. Therefore, the following section first describes autonomous cooperation in general, and then describes the intelligent container and its features as one example of autonomous cooperation-based technologies.
3.3 Autonomous Cooperation-Based Technologies Autonomous cooperation rests upon the concept of self-organization [12], which is researched in many natural scientific disciplines. Some of this natural scientific research on self-organization is on topics such as dissipative structures [29], chaos theory [30], synergetics [31,32], and autopoiesis [33]. Also, in other scientific fields such as logistics (e.g. [34]), researchers investigate phenomena, implications, and possibilities of self-organization specifically autonomous cooperation. The idea behind using the concept of autonomous cooperation for logistics systems is to obtain “a flexible self-organizing system structure which is able to cope with dynamics and complexity while maintaining a stable status.” (Windt and Hülsmann [12], p. 8). Autonomous cooperation describes processes of decentralized decision-making in heterarchical logistics systems [12]. Autonomous logistics objects (e.g. packaged goods, transport systems) can interact with each other and are able to make decentralized decisions independently from external entities. This leads to the non-predetermined behavior of the whole system [35,36]. Therefore, the degree of autonomous cooperation increases as each of these features (decentralized decision-making, heterarchy, interaction, non-predeterminism, and autonomy) is
3
Creating Customer Value in Logistics: Contributions and Limitations
19
gradually increased. How can this theoretical concept of autonomous cooperation be realized in supply chain management? Realization becomes possible through developments in information and communication technologies such as RFID, GPS, UMTS, wireless sensor networks etc. [14,37,37–39]. If these technologies are implemented, for example, into a container, this container can become an ‘intelligent’ logistic entity [40, 41] – what Wycisk et al. [42] call a ‘smart part’ acting in a Complex Adaptive Logistics System. These smart parts are enabled to process information, and to take and execute decisions [42]. This becomes possible through RFID for identification, GPS for locating, and sensor networks to obtain relevant data (from inside the container about e.g. temperature as well as data from outside the container about traffic etc.), and software agents for autonomous decision making [13]. Software agents are local decisionmakers, which are assisted by means of communication, sensors, and calculating capacity [40]. Building an intelligent container by implementing these technologies offers certain features which could be used by a logistics service provider to adapt the offered services to customer demands in order to create additional customer value. Therefore, these features are described next so as to help firms to decide whether using intelligent containers can be an option for offering new transportation services and, consequently, for creating more customer value. The intelligent container has two major features compared to “normal” containers: first, it offers additional information, e.g. on transport conditions or the condition of the goods inside the container, and second, it enables autonomous reaction based on this information [14, 15, 40]. The first feature of the intelligent container is its ability to provide extra information about various transport circumstances and about the condition of the goods inside the container [15]. On the one hand, the intelligent container can provide real-time information on the goods’ current condition if the transportation process is permanently monitored. If transport circumstances such as temperature, humidity etc. in the container are measured, the container can signal this information to an information base [14]. On the other hand, based on this real-time information, present and future quality conditions of the transported goods (such as the ripening of fruit) can be calculated with respect to special points in time (e.g. arrival time and/or expected shelf life) [43]. The second feature of the intelligent container is its ability to react autonomously to critical events based on previously defined instructions and certain information from inside as well as outside the container (therefore, the first feature “real-time information” is essential for the second feature). For example, the container can contact a routing program or transport operator [14,41,44] if, for example, the shelflife forecast falls below a defined value or if the temperature unexpectedly rises due to a broken cooling element. If necessary, the container can be redirected to a closer or less demanding customer or get priority transportation (e.g. airplane instead of vessel) [45] and loss of the container’s content can be avoided. For the future, it might be possible for containers to communicate with each other while taking external factors (e.g. demand forecasts) into account so as to find the
20
M. Hülsmann et al.
most efficient solution by interacting and deciding autonomously [14]. However, the intelligent container can also contact a central authority entity [45] which implies more interaction than a standard container is capable of. In fact, a container could contact a human authority or an entity along the lines of the neural network computational model described by McKelvey et al. [46]. Moreover, ‘smart’ containers can make decentralized, autonomous decisions. Its behavior also is not predetermined, since an intelligent container follows certain rules that are previously defined, but the choices of a particular container are not. While an intelligent container is not yet a fully ‘smart part’ – as it, for example, can not interact with other containers – it still increases the degree of autonomous cooperation in logistics [41]. Utilizing both features (ability to provide significant and real-time information and ability to react autonomously), a logistics service provider might be able to offer new or improved services and consequently better achieve the earlier mentioned seven R’s. Thus, the question then arises: Can a logistics service provider successfully differentiate his/her firm from its competitors by using intelligent containers embedded in an autonomos cooperation-based technology? By way of responding to this question, our next section discusses contributions and limitations facing logistics service providers aiming to create value for their customers by implementing intelligent-container logistics technology.
3.4 Analyzing Positive and Negative Effects of the Intelligent Container on Customer Value Creation This section offers an example analysis of positive and negative implications on the customer value created by implementing an intelligent container (as autonomous cooperation-based technology) for transportation services. This will be done by discussing the implications of offering services based on an intelligent container’s features (as described in Sect. 3.3) on customer value dimensions (as explained in Sect. 3.2: utility value, psychic value, financial costs, and non-financial costs). For a short overview, some selected examples are given in Table 3.1. The ability to provide real-time information about the goods’ condition inside a container, as well as about the goods’ expected shelf life, can increase the customer’s benefits and especially the utility value for the logistics service provider’s customer, since more transparency about the goods’ condition can be generated (most likely due to better access to the right information). Furthermore, this helps to better achieve the goal of the right amount of the right product at the right place at the right time as information about the goods’ position and condition is available in real-time. This means that the logistics service provider can increase his/her customers’ process reliability, which is due to better transparency and more reliable information. However, utility value can also be decreased if too much information is available that cannot be analyzed and thus overloads the information processing capacity (for information capacity overload, see e.g. Probst (2006)). Thus, the right amount of information needs to be found. Also, the second feature – ability to react
3
Creating Customer Value in Logistics: Contributions and Limitations
21
Real-time information
Autonomous reaction
Intelligent container features
Table 3.1 Example effects of implementing the intelligent container on customer value dimensions Customer value dimension Benefits Costs Utility value Psychic value Financial costs Non-financial cost C More Image to be Increased Increased effort transparency reliable transportation of information cost processing Danger of information overload C Increased flexibility
Data security Decreased issues storage cost Image to be innovative
Danger of Fear of loss locally of control optimal solutions that are globally suboptimal
Less quality inspections of goods necessary Increased More effort for costs programming the container Less wastage Learning effects of perishable goods
autonomously to critical events – can contribute to the customers’ benefits amongst others by increasing the utility value. Based on information about the condition of the goods inside the container, quick and flexible changes of routes, destinations, etc., during transportation become possible [14, 40]. For example, there could be delays due to traffic jams or shorted shelf life due to a broken cooling aggregate, which would be noticed because the temperature in the container rises. This can lead to better achievement of the right product at the right place at the right time in the right condition and thus increase the utility value for the customer. However, an increased flexibility to change routes or destinations also involves the risk of locally optimal solutions, which might be globally suboptimal [46]. The ability to provide real-time information also positively affects the customers’ psychic value as e.g. the logistics service provider’s customers can – based on their knowledge about their products’ transport conditions – advertise their special reliability or guarantee a certain quality (long shelf life, low carbon footprint [47]) which can increase their image. Opposed to a reliable positive image, the logistics service provider’s customer might have some fears for the security of his/her data due to an increased transparency. The ability to react autonomously can also provide psychic value to customers as it can support their image of being very innovative. Customers, however, might also fear some loss of control – and therefore loss of psychic value – if containers can react autonomously without consulting an external entity. However, the value aspects also may affect customers’ costs. On the one hand, financial costs probably increase because the intelligent container’s development and investment costs (as more equipment is needed) are higher. Therefore, transportation costs for customers may also increase. On the other hand, storage costs
22
M. Hülsmann et al.
might decrease as replenishment can be forecasted more reliably and less safety stock is necessary. Furthermore, the second feature of the intelligent container (autonomous response capability) can also lead to certain sacrifices for a customer, which can decrease the customer’s received value. For example, transport costs (i.e. financial costs) can be increased via autonomous reaction if, for example, the transport mode is changed from vessel to airplane due to shortened shelf life. But, cost can also decrease as less waste incurs [48], because products with shorter shelf life can (autonomously) more quickly be re-routed with a higher priority so as to reach their destination with an acceptable level of quality, which would not necessarily be possible without autonomous reaction [41]. Non-financial costs may increase because more information leads to an increase of required information processing. However, non-financial costs can also decrease if the effort for quality inspections becomes unnecessary or can be reduced, as the goods’ condition can be more reliably estimated. The ability to react autonomously also affects a customer’s non-financial costs, such as effort needed to define and program the reaction rules for an intelligent container, depending on the kind of goods transported and reactions desired. However, if the intelligent container possesses adequate learning abilities, it can improve its behavior and, therefore, further reduce the non-financial costs. In summary, our discussion shows that implementing intelligent containers for transportation services affects value for logistics service providers’ customers positively [49] (e.g. through more transparency about the transportation conditions of the goods or a better due date reliability by autonomously adapting the transport route), as well as negatively (e.g. an increase of the transport price through the necessity to acquire additional equipment in order to run the transport via the use of the intelligent container). This implies that use of intelligent containers can lead to value improvements as well as losses regarding the creation of customer value. Therefore, net customer value (difference between customer value and customer costs) cannot be assumed as positive for sure. Although this research could gain the above-discussed results, some qualifications exist. First, accurate customer value effects could not be measured. This results from the fact that no information regarding the perception of value by the customers or, more specifically, their individual preferences was available. Second, the strength of effects that result from an implementation of an intelligent container could not be measured as well. An exact measurement of the presumed effects would require empirical research and the creation of a model, which could then show the causal interrelations between the features of the intelligent container and the effects on customer value. Therefore, access to specific customer preference data would be necessary. Third, these effects are exemplarily enumerated for the intelligent container. Therefore, additional effects are possible when implementing the intelligent container methods, and this discussion does not necessarily allow generalizing the effects of autonomous cooperation-based technologies, in general, to all industries or products. However, although further research is required, it can be said, based on the discussion above, that autonomous cooperation-based technologies offer potentials for creating
3
Creating Customer Value in Logistics: Contributions and Limitations
23
customer value, and therewith for logistics service providers to differentiate from competitors.
3.5 Conclusions and Outlook Logistics Service Providers, which act in highly competitive markets, have to cope with a number of challenges because of the homogeneity of services offered by logistics actors. In order to cope with this challenge of service homogeneity, logistics service providers have to differentiate from competitors by developing new or adopt existing services which allows them to create additional customer value. One approach to realize higher value for customers stems from using autonomous cooperation-based technologies in order to create new or to improve existing logistics services. Positive as well as negative effects of implementing the intelligent container as an example for autonomous cooperation-based technologies can be identified that generate or reduce customer value for logistics service providers. Examples of positive effects are more transparency of the goods’ condition during transport and less wastage of perishable goods; examples of negative effects are an increased danger of information overload and the risk of locally-optimal solutions which may be globally suboptimal. All in all, the net effect on customer value cannot in general be assumed positive. These ambiguous possible results lead to the need of further (especially empirical) research in order to determine the net effect on the customer value. However, this analysis does show that autonomous cooperation-based technologies can, in principle, be used for logistics service providers so as to differentiate themselves from competitors, and should therefore be seriously considered when designing new or additional services.
References 1. Dornier P, Ernst R, Fender M, Kouvelis P (1998) Global operations and logistics – text and cases. Wiley, New York 2. Christopher M (2005) Logistics and supply chain management: creating value-adding networks. Financial Times/Prentice Hall, Harlow [u.a.] 3. Person G, Virum H (2001) Growth strategies for logistics service providers: a case study. Int J Logist Manag 12(1):53–64 4. Zhang Q, Vonderembse MA, Lim J (2005) Logistics flexibility and its impact on customer satisfaction. Int J Logist Manag 16(1):71–95 5. Davis DF, Golicic SL, Marquardt AJ (2008) Branding a B2B service: does a brand differentiate a logistics service provider? Ind Market Manag 37:218–227 6. Ulaga W, Chacour S (2001) Measuring customer-perceived value in business markets: a prerequisite for marketing strategy development and implementation. Ind Market Manag 8 30(6):525–540
24
M. Hülsmann et al.
7. Khalifa AS (2004) Customer value: a review of recent literature and an integrative configuration. Manag Decis 05 42(5):645–666 8. Lam SY, Shankar V, Erramilli MK, Murthy B (2004) Customer value, satisfaction, loyalty, and switching costs: an illustration from a business-to-business service context. J Acad Market Sci Summer2004 32(3):293–311 9. Butz HE, Goodstein LD (1996) Measuring customer value: gaining the strategic advantage. Organ Dyn 24:63–77 10. Woodruff RB (1997) Customer value: the next source for competitive advantage. J Acad Market Sci 25(2):139–153 11. Flint DJ, Larsson E, Gammelgaard B, Mentzer JT (2005) Logistics innovation: a customer value-oriented social process. J Bus Logist 03 26(1):113–147 12. Windt K, Hülsmann M (2007) Changing paradigms in logistics – understanding the shift from conventional control to autonomous cooperation and control. In: Hülsmann M, Windt K (eds) Understanding autonomous cooperation & control – the impact of autonomy on management, information, communication, and material flow. Springer, Berlin. p. 1–16 13. Gehrke JD, Behrens C, Jedermann R, Morales-Kluge E (2006) The intelligent container – toward autonomous logistic processes KI 2006 Demo Presentations, p. 15–18 14. Jedermann R, Behrens C, Laur R, Lang W (2007) Intelligent containers and sensor networks, approaches to apply autonomous cooperation on systems with limited resources. In: Hülsmann M, Windt K (eds) Understanding autonomous cooperation & control in logistics – the impact on management, information and communication and material flow, p. 365—392. Springer, Berlin 15. Jedermann R, Lang W (2009) In: Rödig U, Sreenan CJ (eds) The minimum number of sensors – interpolation of spatial temperature profiles, p. 232–246. Springer, Berlin/Heidelberg 16. Narasimhan R, Das A (1999) An empirical investigation of the contributions of strategic sourcing to manufactering flexibilities and performance. Decis Sci 30(3):683–718 17. Juga J, Pekkarinen S, Kilpala H (2008) Strategic positioning of logistics service providers. Int J Logist Res Appl 12 11(6):443–455 18. Porter ME (1985) Technology and competitive advantages. J Bus Strat 19. Day GS (1990) Market driven strategy. Free Press, New York 20. Teece DJ (1990) Contributions and impediments of economic analysis to the study of strategic management. In: Fredrickson JW (ed) Perspectives on strategic management, p. 39–80. New York 21. Richey RG, Daugherty PJ, Roath AS (2007) Firm technological readiness and complementarity: capabilities impacting logistics service competency and performance. J Bus Logist 28(1):195–228 22. Grönroos C (1997) Value-driven relational marketing: from products to resources and competencies. J Market Manag 13(5):407–419 23. Lapierre J (2000) Customer-perceived value in industrial contexts. J Bus Indust Market 15(2/3):122–140 24. Zeithaml V (1988) Consumer perception of price, quality and value: a means-end model and synthesis of evidence. J Market 52:2–22 25. Huber F, Herrmann A, Morgan RE (2001) Gaining competitive advantage through customer value oriented management. J Consum Market 18(1):41–53 26. Parolini C (1999) The value net: a tool for competitive strategy. Wiley, Chichester 27. Groth JC (1994) The exclusive value principle: a concept for marketing. J Product Brand Manag 3(3):8–18 28. Mentzer JT, Flint DJ, Hult TM (2001) Logistics service quality as a segment-customized process. J Market 65(4):82–104 29. Prigogine I (1969) Structure, dissipation and life. In: Marois M (ed) Theoretical physics and biology, p. 23—52. North-Holland Publishing, Amsterdam, London 30. Peitgen H (1986) The beauty of fractals: images of complex dynamical systems. Springer, Berlin [u.a.] 31. Haken H (1977) Synergetics: a workshop; proceedings of the international workshop on synergetics at schloss elmau, bavaria, may 2–7, 1977. Springer, Berlin [u.a.]
3
Creating Customer Value in Logistics: Contributions and Limitations
25
32. Foerster VH (1960) On self-organizing systems and their environment. In: Yovits MC, Cameron S (eds) Self-organizing systems, p. 31–50. Pergamon, London 33. Maturana HR, Beer S (1980) Autopoiesis and cognition: the realization of the living. Reidel, Dordrecht [u.a.] 34. Hülsmann M, Windt K (eds) Understanding autonomous cooperation & control – the impact of autonomy on management, information, communication, and material flow, 1st edn. Springer, Berlin 35. Hülsmann M, Grapp J, Li Y (2008) Strategic adaptivity in global supply chains – competitive advantage by autonomous cooperation. Int J Prod Econ 7 114(1):14–26 36. Hülsmann M, Grapp J (2005) Autonomous cooperation in international-supply-networks – the need for a shift from centralized planning to decentralized decision making in logistic processes. In: Pawar KSea (ed) Proceedings of the 10th international symposium on logistics (10th ISL), p. 243–249. Loughborough, UK 37. Böse F, Windt K (2007) Catalogue of criteria for autonomous control in logistics. In: Hülsmann M, Windt K (eds) Understanding autonomous cooperation and control in logistics, p. 57–72, 1st edn. Springer, Berlin 38. Scholz-Reiter B, Windt K, Freitag M (2004) Autonomous logistic processes: new demands and first approaches. In: Monostori L (ed) Proceedings of the 37th CIRP international seminar on manufacturing systems, Budapest, p. 357–362 39. Philipp T, de Beer C, Windt K, Scholz-Reiter B (2007) Evaluation of autonomous logistic processes – analysis of the influence of structural complexity. In: Hülsmann M, Windt K (eds) Understanding autonomous cooperation & control in logistics – the impact on management, information and communication and material flow, ; p. 303–324, 1st edn. Springer, Berlin 40. Jedermann R, Behrens C, Westphal D, Lang W (2006) Applying autonomous sensor systems in logistics – combining sensor networks, RFIDs and software agents. Sensor Actuator A: Phys 11/8 132(1):370–375 41. Hülsmann M, Lang W, Colmorn R, Jedermann R, Illigen C, Cordes P (2010) In: Pawar KS, Lalwani CS (eds) Ecological impacts of autonomously co-operating technologies in international supply networks – establishing a model for analyzing CO2-effects of implementing intelligent containers in fruit logistics. Kuala Lumpur, Malaysia 42. Wycisk C, McKelvey B, Hülsmann M (2008) “Smart parts” supply networks as complex adaptive systems: analysis and implications. Int J Phys Distribut Logist Manag 38(2):108–125 43. Tijskens LMM, Polderdijk JJ (1996) A generic model for keeping quality of vegetable produce during storage and distribution. Agr Syst 8 51(4):431–452 44. Jedermann R, Gehrke JD, Lorenz M, Herzog O, Lang W (2006) Realisierung lokaler Selbststeuerung in Echtzeit: Der Übergang zum intelligenten Container. In: Pfohl HC, Wimmer T (eds) Wissenschaft und praxis im dialog. Steuerung von logistiksystemen – auf dem weg zur selbststeuerung, p. 145–166. Deutscher Verkehrs-Verlag, Hamburg 45. Jedermann R, Schouten R, Sklorz A, Lang W, van Kooten O (2006) Linking keeping quality models and sensor systems to an autonomous transport supervision system. In: Kreyenschmidt J, Petersen B (eds) Cold chain-management. Proceedings of the 2nd international workshop cold chain management, p. 3–18. University Bonn, Bonn 46. McKelvey B, Wycisk C, Hülsmann M (2009) Designing an electronic auction market for complex ‘smart parts’ logistics: options based on LeBaron’s computational stock market. Int J Prod Econ 08 120(2):476–494 47. Deurer M, Clothier B, Pickering A (2008) How will carbon footprinting address the issues of reduction, mitigation, emissions trading and marketing 48. Jedermann R, Behrens C, Lang W (2006) Mit Agenten der Frische auf der Spur. Frischelogistik, Fachmagazin für die gesamte Frische- und Tiefkühlkette 6:22–25 49. Hülsmann M, Grapp J (2008) Economic success of autonomous cooperation in international supply networks? – designing an integrated concept of business modelling and service engineering for strategic usage of transponder-technologies. In: Pawar KS, Lalwani CS, Banomyong R (eds) Integrating the global supply chain. Conference proceedings of 13th international symposium on logistics (13th ISL), p. 117–124. Loughborough, UK
Chapter 4
Autonomous and Central Control of Production Networks Sergey Dashkovskiy, Andrii Mironchenko, and Lars Naujok
4.1 Introduction Production networks are typical examples of large-scale and complex systems with a nonlinear behavior. By an increasing number of entities and material within the network the control of the material flows and the production rates of entities or machines is not an easy task. One way to achieve logistic goals, such as a high performance, robustness and stability, is the shift from centralized to decentralized or autonomous control. The term production network is used to describe company or cross-company owned networks with geographically dispersed plants. These types of networks may react quickly on perturbations due to redundancies of common resources. But high flexibility causes interdependencies between production processes in different plants, e.g., allocation problems for products or planning of transports and transport capacity [3, 16]. Therefore production planning and control (PPC) of production networks has to cover these tasks and also has to provide methods for an integrated planning and synchronization within the network, including planning of sales and inventory [28]. Under highly dynamic and complex conditions current PPC methods cannot cope with disturbances or unforeseen events in an appropriate manner [15]. Changing market conditions and inappropriate planning may cause uncertainties of lead times, inconstancy of schedules or may also lead to instability or even chaos. The main idea of autonomous cooperating logistic processes is to enable intelligent logistic objects to route themselves through a logistic network according to their own objectives and to make and execute decisions, based on local information [29, 30]. In this context intelligent logistic objects may be physical or material objects, e.g., parts or machines, as well as nonmaterial objects (e.g., production orders, information). It has been already shown that different autonomous control S. Dashkovskiy (B), A. Mironchenko, and L. Naujok ZeTeM, Universität Bremen, Bibliothekstr.1, 28359 Bremen, Germany e-mail:
[email protected],
[email protected],
[email protected] M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_4,
27
28
S. Dashkovskiy et al.
methods can help to increase the logistics performance and robustness of single production systems [17–19]. On the other hand, autonomously controlled production networks may show a sudden change of the dynamical systems behavior in dependence of varying start parameters and the logistic performance collapses in the sense of unpredictable and increasing throughput times and growing inventory [21]. Thus, investigations of the stability of autonomously controlled production networks are essential to understand the dynamical systems behavior. Typical examples of unstable behavior are unbounded growth of unsatisfied orders or unbounded growth of the queue of the workload to be processed by a machine. This causes high inventory costs and loss of customers. To avoid instability of a network it is worth to investigate its behavior in advance. In particular, mathematical modeling and analysis provide helpful tools for design, optimization and control of such networks and for deeper understanding of their dynamical properties. Roughly speaking, for production networks stability means that the state of the network remains bounded over time under bounded external inputs. In this contribution we identify the state as the number of unprocessed parts, which is the sum of the queue length and the work in progress (WIP). To identify parameters which guarantee stability of the network we are going to apply tools from mathematical systems theory. In this context mathematical models describing production network’s behavior are needed. In this contribution we consider the recent results in modeling of logistic networks, based on the input-to-state stability concepts, and analyze, in what form the limitations of control methods appear in the mathematical models of logistic processes. We argue, that in case, when the extensive information about the system is available, the central control may be quite effective, and can provide at least some important information about the structure of controls. On the other hand, if there is only a little information about the network, the autonomous control is a good alternative, but it rises the question about the stability of the network, which can be answered with the help of mathematical stability theory. We discuss these results from the viewpoint of limitations of autonomous control, and provide a number of parameters, that can be used to find the bounds, within which autonomous control is more effective, than central planning. The structure of the contribution is as follows. In Sect. 4.2.1 we model general production network, which we analyze by means of optimal control methods in Sect. 4.2.2. In Sect. 4.2.3 we review recent results in autonomous control of production networks. Then we identify parameters of the network which influence the comparative effectiveness of different control methods in Sect. 4.2.4. The notions of stability and the tools to check whether a system is stable are presented in Sect. 4.3. In Sect. 4.4 we identify and discuss the limitations of autonomous control. Finally, the conclusions are given in Sect. 4.5. Throughout the paper by x T we denote the transposition of a vector x 2 Rn ; n 2 N and RnC denotes the positive orthant fx 2 Rn W x 0g where we use the standard partial order for x; y 2 Rn given by x y , xi yi ; i D 1; : : : ; n and x 6 y , 9i W xi < yi :
4
Autonomous and Central Control of Production Networks
29
4.2 Modeling and Control of Production Networks 4.2.1 Description and Modeling of a General Production Network We are going to construct a model of a production network, consisting of n market entities (n is an arbitrary positive number), which may be, for example, raw material suppliers (e.g., extracting or agricultural companies), producers, distributors and consumers. Each entity is a subsystem of the whole network. For simplicity we assume, that there is only one unified type of material, i.e., all primary products, used in the production network, can be measured as a number of units of this unified material (Fig. 4.1). We assume, that i -th subsystem is characterized at time t 2 RC only by the parameter xi .t/, which is the quantity of unprocessed material within the i -th subsystem at time t. The state of the whole network is denoted by x.t/ D .x1 .t/; : : : ; xn .t//T . A subsystem can get material from an external source, which is denoted by ui (external inputs), and from subsystems of the network (internal inputs). Let the i -th subsystem processes the raw material from its inventory with the rate fQi i .t; x.t// 0 and sends the produced goods (measured in units of unified material) to the j -th subsystem with the rate fQj i .t; x.t//. Thus, the total rate of the P distribution from the i -th subsystem to other subsystems is njD1 fQj i .t; x.t//. It may send a part of production to the customers not considered in the network. In the version of the model, that we consider in this contribution, we do not distinct the transport of the production out of the network and the losses of production in the manufacture process. However, one can slightly generalize the model, to include the systems with outputs, where output is the total flow of goods out of the system. In this case it is useful to use instead of input-to-state stability (see Sect. 4.3) the notion of input-to-output stability. For general functions fQj i it is hard to derive stability conditions. Therefore we consider the following case: fQj i .t; x.t// D cj i .x.t//fQi .xi .t//; cj i .x/ 2 RC 8x 2 RnC and fQi i .t; x.t// D cQi i .x.t//fQi .xi .t//; cQi i .x/ 2 RC ; x 2 RnC , where fQi 2 K
Fig. 4.1 Example of supply network
30
S. Dashkovskiy et al.
is proportional to the processing rate of the system, cj i .x.t//; i ¤ j are some positive distribution coefficients and cQi i .x.t// 0. We will denote them cj i .t/ for the sake of brevity. Under these assumptions the dynamics of the i -th subsystem is described by ordinary differential equations: xP i .t/ D
n X
cij .t/fQj .xj .t// C ui .t/ cQi i .t/fQi .xi .t//; i D 1; : : : ; n: (4.1)
j D1;j ¤i
Denoting ci i WD cQi i we can rewrite the above equations as an interconnected system in a vector form x.t/ P D C.t/fQ.x.t// C u.t/;
(4.2)
where fQ.x.t// D .fQ1 .x1 .t//; : : : ; fQn .xn .t///T , u.t/ D .u1 .t/; : : : ; un .t//T and C.t/ 2 Rnn . The distribution coefficients may be used as controls for some central or autonomous control method. If the elements of the subsystems are controlled autonomously, then every subsystem can use their own control method, and it is important to have conditions, that guarantee, that the system is stable. If the system is controlled by central planning method, then we assume, that there exists a planning center, that chooses the distribution coefficients to reach some aim, for example, to maximize the total production rate of the network, to minimize the transportation costs, to improve the stability properties of the system. In the following we show, how one can use both methods to obtain the stability conditions and gain some information, how to construct the possible distributional coefficients.
4.2.2 Optimal Strategies In this section we model the planning center of a production network (4.1) as an object, that controls directly all distributional coefficients cij , i; j D1; : : : ; n; i ¤ j . That is, we assume in this section, that cij are functions only of t, and do not depend on x.t/ explicitly. For convenience we assume, that the aim of the planing center is to minimize the costs for the storage of the material over the time Œ0; T , which can be modeled by the following functional: Z S1 D
0
T
n X
xi2 .s/ds ! mi n:
(4.3)
i D1
To achieve this goal it uses the information about the inputs ui , i D 1; : : : ; n (which is known for some time interval Œ0; T ).
4
Autonomous and Central Control of Production Networks
31
The controls, that a planning center may use to achieve this aim, are the distributional coefficients cij .t/ 0, i; j D 1; : : : ; n, i ¤ j . We assume, that there are given the following constraints to the distributional coefficients: n X
cij .t/ D kj ; j D 1; : : : ; n:
(4.4)
i D1;i ¤j
They tell us, that each node of a network can control only the portions of goods, distributed to other nodes, but cannot control the total distribution rate. Our aim is to prove, that under made assumptions the functions cij , i; j D 1; : : : ; n; i ¤ j have to be, for realistic inputs, piecewise constant functions. The necessary conditions of extremum are given by the Pontryagin Maximum Principle. The Hamiltonian for system (4.1) with objective functional (4.3) is given by: HD
n X
0 pi .t/ @ci i .t/fi .xi .t// C
i D1
1
n X
cij .t/fj .xj .t// C ui .t/A0
n X
xi2 .t/:
i D1
j D1;j ¤i
(4.5) Here pi , i D 1; : : : n and 0 are the Lagrange coefficients. According to Pontryagin Maximum Principle (see, e.g., [1]), cij 0 have to be chosen such, that condition (4.4) holds, and H ! max. We cannot control inputs ui and consumption rate ci i , i D 1; : : : ; n, and therefore maximization of (4.5) leads us to the following problem: n X
0 pi .t/ @
i D1
D
1
n X
cij .t/fj .xj .t//A D
j D1
0 @
n X
n X
pi .t/cij .t/fj .xj .t//
i D1 j D1;j ¤i
j D1;j ¤i n X
n X
1 pi .t/cij .t/A fj .xj .t// ! max:
i D1;i ¤j
To maximize this linear combination of fj .xj .t//, which are nonnegative for all t 0. The coefficients cij are dependent only on the distributional coefficients from the same sum in brackets, therefore the terms in brackets may be maximized independently 8j D 1; : : : ; n
n X
pi .t/cij .t/ ! max:
i D1;i ¤j
Using constraints, we conclude, in case, if the set fp1 .t/; : : : ; pn .t/g has the unique maximal element, that
32
S. Dashkovskiy et al.
8j D 1; : : : ; n; 8s D 1; : : : ; n; s ¤ j; csj .t/ kj ; s D arg maxfp1 .t/; : : : ; pn .t/g D 0; otherwise If for some t there exists I f1; : : : ; ng, such that 8i; j 2 I it holds, that pi D pj and 8i 2 I; 8k 2 f1; : : : ; ng pi pk . Then 8j D 1; : : : ; n; 8s D 1; : : : ; n; s ¤ j; csj .t/ 2 Œ0; kj ; s 2 I D 0; otherwise and it must still hold (4.4). These controls are called mixed controls. But the situation, that the values of at least two functions pi , i D 1; : : : ; n will be identical on some time interval Œt0 ; t1 , t1 > t0 is quite unrealistic, therefore we see, that the controls cij are (with exception to quite unrealistic inputs) piecewise constant functions. One can interpret it, that in every moment one has to send the goods to the node, that “promises” the biggest virtue.
4.2.3 Autonomous Control Methods In this section we describe several autonomous control policies. They can be implemented in the model of a production network by the terms cij .t/ in (4.1). These different policies were developed and investigated in the research project CRC637 (see the Acknowledgement). The queue length estimator (QLE) policy enables logistic objects in a production network to estimate the waiting, processing and transportation times of different alternative processing resources. A logistic object will decide for the plant with the lowest estimated waiting, processing and transportation time. It uses local information to evaluate the states of the alternatives. It was shown in [26] that the application of this policy leads to a better systems performance regarding throughput times compared to classical scheduling algorithms in highly dynamic situations. In the model (4.1) the distribution rates of the QLE policy can be defined, for example, by cij .t/ WD cijq .t/ and cijq .t/ WD
1 xi .t /C" P 1 k x .t /C" k
;
where the index k denotes all subsystems which get material from subsystem j and " > 0, arbitrarily small, is inserted to let the fraction be well-defined. The interpretation of cijq is in simple words the following: if the queue length of the i th subsystem is small, then more material will be sent to subsystem i in contrast to the case where xi is large and cijq is small.
4
Autonomous and Central Control of Production Networks
33
Similar to the QLE, the due date (DUE) policy estimates waiting, processing and transportation times. While the QLE uses this information for minimizing throughput times of logistic objects, the DUE policy orientates at the tardiness of logistic objects. A logistic object using this policy decides for an alternative resource which offers the lowest difference between estimated due date and pre-planned due date [22]. The pheromone based policy is a bio-inspired strategy. The approach is based on the idea to imitate the process of marking possible routes to food sources by ants. Ants emit pheromones between the nest and food sources. Other ants can detect those pheromones and will follow the trail with the highest concentration of pheromones. This is transferred to production networks: During the production process, the logistic objects leave information about their transportation, processing and waiting times at a corresponding machine or plant. Following logistic objects compare this artificial pheromone concentration and choose a production line. Thus, the pheromone concentration depends on transportation, waiting and processing times of previous logistic objects [4, 25]. To model the evaporation process of natural pheromones a constant vi 2 RC is inserted in the definition of the distribution rates, which can be chosen, for example, as X fQi .xi .t// fQk .xk .t// C ; vk P Q Qq .xq .t// C " f k fk .xk .t// C " q k¤i
cij .t/ D cijp .t/ WD .1 vi / P
where k; q are indices denoting the subsystems which get material from subsystem j, " > 0 and 0 vi 1 is the evaporation constant of the i th subsystem. By the evaporation constant vi one can justify the PHE method in order to increase the performance or robustness of the network. The honey bee algorithm (HBA) is another bio-inspired strategy. It uses the foraging mechanisms of honey bees’ colonies. In nature, bees advertise possible food sources with a so called ‘waggle dance’. The duration of this dance depends on the ratio between energy consumption of the flight (between hive and food source) and available energy of the source. The probability of bees recognizing the dance of a dancing bee is proportional to the dancing duration. According to this principle logistic objects are able to advertise different alternative production resources by means of the machining quality, which is determined by calculation of the ratio of value added and the throughput time needed for this step [20]. The natural process, which inspires the chemotaxis (CHE) policy, differs from the PHE and the HBA policy. It is not inspired by coordination principals of social insects, but on movement processes coming from microbiology. Natural bacteria are able to direct their movement according to the concentration of attractants (e.g., food substances) or repellants (e.g., toxic substances). Therefore, bacteria perform a random biased walk to find appropriate food sources. This basic movement principle is transferred to autonomous decision making by the CHE policy. Logistic objects using this policy decide according to the gradient of logistic target values of different decision alternatives [24].
34
S. Dashkovskiy et al.
In the following section we introduce a couple of parameters that characterize the information distribution throughout the network.
4.2.4 Properties of a Network As we will see, neither autonomous control nor central planning is a panacea for solving of all problems, that arise in the control of production networks. The central planning, if it is theoretically and practically possible, may provide better results, than autonomous control. But usually the problems, that arise in real networks are so complex, that the decisions, made by the central planning method will take a lot of resources (e.g., too much computation time), and the state of the systems changes so quickly, that the obtained results will have in a new situation only limited applicability. Another problem, that arises in supply networks, is the lack of information in the network, when it is sent from the subsystems of the network to the planning center. Managers, that make decisions for the subsystems have always more information about their own systems, than they send to the planning center. Consequently, although the planning center has information about all the nodes of the network, this information is not exact, and decisions of a planning center may be not accurate enough or at all harmful for the system. In this section we introduce three properties, that allow to choose the appropriate method to control a production network: n - Size of a network. p - A number, that characterizes availability of information about the network. L - Loss of information in hierarchical structure of a network.
In networks of small size the central control can provide better results, because the complexity of the problem is lower than for large scale networks, and central control can use the knowledge of the state of all the system more effectively. Availability of information in a production network plays also a big role. The more knowledge has the planning center, the more effective will be the work of control center, and the less information is given, the better results will be provided by the autonomous control strategy. We collect these considerations to the following table:
p is small p is large
n is small Both Central planning
n is large Autonomous control Mixed strategy
With the increase of L, the advantages of the autonomous control over central planning also increase. For modeling of autonomous control it is important to investigate the stability of the network. We have to be sure, that the decisions, made by subsystems without
4
Autonomous and Central Control of Production Networks
35
(or almost without) communication with each other, cannot induce the unstable behavior of the system. Since stability is important for the performance and vitality of the network we provide a framework, that may be used to gain some information about stability of the network, and we state the recent results about input-to-state stability of the model of production network, shown in the Sect. 4.2.1.
4.3 Stability of Production Networks In this section we recall stability properties and tools how to check whether a system is stable. One possibility to model production networks are ordinary differential equations (ODEs). An ODE is of the form x.t/ P D f .x.t/; u.t//; t 2 RC ;
(4.6)
where x 2 RN denotes the state of the system, u 2 RM is the essentially bounded measurable external input and f W RN RM ! RN describes the system dynamics. ODEs describe the evolution of the state of the system with continuous time t 2 RC , where RC WD Œ0; 1/. To have existence and uniqueness of a solution of a system of the form (4.6) the function f is assumed to be a locally Lipschitz continuous function. The solution is denoted by x.tI x0 ; u/ or x.t/ for short, where x0 WD x.0/ is the initial condition. In general, production networks consist of n 2 N interconnected systems of the form xP i .t/ D fi .x1 .t/; : : : ; xn .t/; ui .t//; t 2 RC ; i D 1; : : : ; n;
(4.7)
Pn
where xi 2 RNi , ui 2 RMi and fi W R j D1 Nj CMi ! RNi are locally Lipschitz continuous functions. Here, xj ; j ¤ i can be interpreted as internal inputs of the i -th subsystem and the solution is denoted by xi .tI xi0 ; xj ; j ¤ i; ui / or xi .t/ for short, where xi0 WD xi .0/Pis the initial condition. Pn n T T T If we define N WD i D1 Ni ; M WD i D1 Mi , x WD .x1 ; : : : ; xn / , u WD T T T T T T .u1 ; : : : ; un / and f D .f1 ; : : : ; fn / , then the interconnected system of the form (4.7) can be written as one single system of the form (4.6), which we call the whole system. The purpose of this section is to analyse production networks, which can be written in the form (4.7), in view of stability. For this purpose we introduce: Definition 1. We define following classes of functions: P WD ff W Rn ! RC j f .0/ D 0; f .x/ > 0; x ¤ 0 g K WD f W RC ! RC j is continuous, .0/ D 0 and strictly increasing g
36
S. Dashkovskiy et al.
K1 WD f 2 K j is unboundedg L WD f W RC ! RC j is continuous and strictly decreasing with o lim .t/ D 0 t !1
KL WD fˇ W RC RC ! RC j ˇ is continuous, ˇ.; t/ 2 K; ˇ.r; / 2 L; 8t; r 0g We call functions of class P positive definite. Definition 2. 1. System (4.6) is locally input-to-state stable (LISS) if there exist constants ; u > 0; 2 K and ˇ 2 KL such that for all initial values jx0 j and all inputs kuk1 u the inequality jx.t/j max fˇ .jx0 j ; t/ ; .kuk1 /g is satisfied 8 t 2 RC , where jj denotes the Euclidean norm and kuk1 WDess sup t 2Œ0;1/
ju.t/j is the essential supremum norm. is called (nonlinear) gain. u 2. The i -th subsystem of (4.7) is called LISS if there exist constants ˇ 0 ˇi ; ij ; i > ˇ ˇ 0; ij ;i 2 K and ˇi 2 KL such that for all initial values xi i and all inputs xj 1 ij ; kui k1 iu the inequality
ˇ ˇ jxi .t/j max ˇi ˇxi0 ˇ ; t ; max ij xj 1 ; i .kui k1 /
j ¤i
is satisfied 8 t 2 RC . ij and i are called (nonlinear) gains. Note that, if ; u D 1 then the system (4.6) is called (global) ISS and if i ; ij ; iu D 1 then the i -th subsystem of (4.7) is called (global) ISS. In particular LISS and ISS guarantee that the norm of the trajectories of each subsystem is bounded. An important tool to verify LISS and ISS, respectively, of a system of the form (4.7) are Lyapunov functions. Definition 3. We assume that for each subsystem of the interconnected system (4.6) there exists a function Vi W RNi ! RC , which is locally Lipschitz continuous, proper and positive definite. Then, for i D 1; : : : ; n the function Vi is called a LISS Lyapunov function of the i -th subsystem of (4.7) if Vi satisfies the following two conditions: There exist functions 1i ; 2i 2 K1 , where K1 is the subset of K-functions that are unbounded, such that 1i
.jxi j/ Vi .xi /
2i
.jxi j/ ; 8 xi 2 RNi
(4.8)
and there exist ij ; i 2 K, a positive definite function i , which is continuous, i .0/ D 0 and i .r/ > 0; 8r 2 R, and constants i ; ij ; iu > 0 such that
4
Autonomous and Central Control of Production Networks
37
Vi .xi / max max ij Vj .xj / ; i .jui j/ ) rVi .xi / fi .x; u/ i .Vi .xi // j ¤i
(4.9) ˇ ˇ ˇ ˇ for almost all xi 2 RNi ; ˇxi0 ˇ i ; ˇxj ˇ ij ; ui 2 RMi ; jui j iu , i i D 0, where r denotes the gradient of the function Vi . Functions ij are called LISS Lyapunov gains. Note that, if i ; ij ; iu D 1 then the LISS Lyapunov function of the i -th subsystem becomes an ISS Lyapunov function of the i -th subsystem (see [14]). In general the LISS Lyapunov gains are different from the gains in Definition 2. Condition (4.8) implies that Vi is proper, positive definite and radially unbounded. Vi can be interpreted as the energy of a system and ˚ the second condition (4.9) of a Lyapunov function means that if Vi .xi / max ˚maxj ¤i ij Vj .xj / ; i .jui j/g holds, then the energy decreases. If Vi .xi / < max maxj ¤i ij Vj .xj / ; i .jui j/g then the energy of the system is bounded by the expression on the left side of the previous inequality. Overall, the trajectory of the system is bounded. Furthermore we define the gain-matrix WD .ij /nn ; i; j D 1; : : : ; n; i i D 0, which defines the map W RnC ! RnC by T .s/ WD max 1j .sj /; : : : ; max nj .sj / ; j
j
s 2 RnC :
(4.10)
Note that the matrix describes in particular the interconnection structure of the network, moreover it contains the information about the mutual influence between the subsystems, which can be used to verify the (L)ISS property of networks. Definition 4. satisfies the local small gain condition (LSGC) on Œ0; w , provided that
.w / < w and .s/ 6 s; 8s 2 0; w ;
s ¤ 0:
(4.11)
Notation 6 means that there is at least one component i 2 f1; : : : ; ng such that .s/i < si . To check whether the interconnected system of the form one has to find a LISS Lyapunov function for each subsystem. If there exists a LISS Lyapunov function for each subsystem then it has the LISS property. Furthermore, if the LISS Lyapunov gains satisfy the local small-gain condition, then the whole system of the form (4.6) is LISS, which we recall in the following theorem (see [13]): Theorem 1. Consider the interconnected system (4.7), where each subsystem has an LISS Lyapunov function Vi . If the corresponding gain-matrix satisfies the local small-gain condition (4.11), then there exist constants ; u > 0, such that the whole system of the form (4.6) is LISS. In [9] a similar ISS small-gain theorem for general networks was proved, where the small-gain condition is of the form
38
S. Dashkovskiy et al.
.s/ 6 s; 8 s 2 RnC n f0g :
(4.12)
Now we use the stability notions and Lyapunov functions for a stability analysis of production networks, modeled in Sect. 4.2.1 in the form (4.1) or (4.2). With help of Theorem 1 we can derive conditions, which guarantee stability of the production network: Proposition 1. Consider a network as in (4.2). 1. Assume that the cij are bounded for all i; j D 1; : : : ; n; i ¤ j and fQj 2 K1 , j D 1; : : : ; n. If 9a 2 Rn ; 2 Rn ; ai > 0; i < 0; i D 1; : : : ; n such that the condition C.t/a < holds 8t > 0, then the whole network is ISS. 2. Let fQj 2 K n K1 , and ˛j WD supxj 2R ffQj .xj /g, j D 1; : : : n; ˛ WD .˛1 ; : : : ; ˛n /T . If 9u 2 L1 .RC ; RnC / and 9 2 Rn ; i < 0; i D 1; : : : ; n such that C.t/˛ C kuk1 < ;
(4.13)
where kuk1 WD .ku1 k1 ; : : : ; kun k1 /T , then the whole network (4.2) is LISS. Furthermore, the constants and u from the Definition 2 may be chosen as WD 1, u WD mini D1;:::;n kui k1 and (4.13) holds for all w 2 L1 .RC ; RnC / W kwk1 kuk1 . The proof can be found in [10]. For certain scenarios of production networks it may happen that the conditions to guarantee stability in Proposition 1 are conservative, which means that one can find parameters which do not satisfy the conditions, but the trajectory of the state of the network is bounded (stable). The reason is, that the presented stability analysis is a “worst-case” approach. To refine the conditions in Proposition 1 and to identify parameters which guarantee stability of a certain scenario of a production network a dual approach of the stability analysis and simulations was presented in [23]: At first, one derives parameters constellations which guarantee stability by the help of Proposition 1. A large set of parameter constellations is checked except the few constellations for which the conditions in Proposition 1 are not satisfied. This set of (stable) parameter constellations is refined by simulation runs of the network, where only the parameter constellations are simulated which do not satisfy the conditions in Proposition 1. In contrast to a pure simulation approach, where the time needed for the identification of parameter constellations which guarantee stability of the network increases exponentially by an increasing number of entities or parts within the network, this dual proceeding identifies parameter constellations in an acceptable time. This procedure was shown and explained in more detail in [23] and can be summarized in the following figure (Fig. 4.2):
4
Autonomous and Central Control of Production Networks
39
Fig. 4.2 Scheme of the identification of (stable) parameter constellations
4.3.1 Possible Generalizations of the Model Naturally, in production networks the time needed for the transportation of material from one plant to another has to be taken into account. To model time-delays (e.g., transportation times) we use functional differential equations of the form x.t/ P D f .xt ; u.t//; t > 0; x. / D 0 . /; 2 Π; 0 ;
(4.14)
where t 2 RC ; x.t/ 2 RN , u.t/ 2 RM is an essentially bounded measurable input and the function xt W Π; 0 ! RN is given byxt . / WD x.t C /; 2 Π; 0, where is the maximum involved delay. f W C Π; 0 I RN RM ! RN is a locally Lipschitz continuous functional on any bounded set. Using interconnections of time-delay systems one can model general production networks with transportation times, where (4.1) or (4.2) must be adapted accordingly (see [10]). Also the ISS property has to be defined for systems of the form (4.14) and interconnected systems of such type. To check whether a system of the form (4.14) has the ISS property, one can use Lyapunov-Razumikhin (LR) functions or Lyapunov-Krasovskii (LK) functionals. It was shown in [8] that if each subsystem of an interconnected system with timedelays admits a LR or a LK function(al) and if the small-gain condition (4.12) is satisfied, then the whole network possesses the ISS property. A certain scenario of an autonomously controlled production network was investigated using the mentioned tools in [10, 11]. Another approach to model production networks is to combine continuous dynamics, described by ODEs or functional differential equations, and discrete dynamics. Such systems are called impulsive or hybrid systems and cover, for
40
S. Dashkovskiy et al.
example, the load and unload of the inventory of a plant. Also time-delays can be considered. The ISS property and the Lyapunov tools have to be defined accordingly. Then, it was shown in [6, 12] that a network of impulsive or hybrid systems possesses the ISS property, if there exists a Lyapunov function for each subsystem and the small-gain condition is satisfied. To include into the model not only transportation times, but the description of the whole transportation process, we may use the partial differential equations (PDE). The simplest model of this art is a standard inhomogeneous transport equation, that takes into account the transportation time and the losses (or increasing) of goods during the transportation. In paper [5] the transport equation was generalized to include the restriction for the traffic flow capacity (in particular, it makes possible to analyze the traffic jams). The input-to-state stability methods for systems, based on partial differential equations, at the time are not developed to the same extent, as the impulsive or time-delay systems. Some preliminary results for certain classes of PDEs were obtained in [7]. To model the central control methods for the logistic networks, one can use the corresponding methods from optimal control of systems with time-delays, systems, based on partial differential equations (see, e.g., [27]) or other classes of systems.
4.4 Limitations of Autonomous Control To find, what are the limitations of autonomous control, we have to answer firstly the question, what are the limitations, common for both central and autonomous control. We distinct: “Essential” limitations, which provide us with bounds, within which the control
method may be applied, both theoretically and in practice. Limitations of effectiveness, which provide us with bounds, within which the
method works better, than the other ones. Returning to parameters of the networks, introduced in Sect. 4.2.4, we can say, that we can identify the following limitations of autonomous control: Structural limitations: The limitations, that are essential for a system, and do not depend on the type of control strategy. In our model such limitations are the stability bounds of the system. It is clear, that if the processing rate of materials is bounded for all xi .t/, then we can always find such an input, that the system will be unstable for all possible controls cij . The condition (4.13) provides us with a priori stability bounds, within which the stability is guaranteed. Limitations of information availability: The less information about the network is available, the less effective is the central control. For example, in retail trade there is no exact information about the quantity of customers, the time of purchases etc., therefore the “pure” central planning, in particular, the direct control of all transportations is not effective, if at all possible. Conversely, for the companies, that
4
Autonomous and Central Control of Production Networks
41
have strict arrangements with other enterprises (in terms of our model, the inputs are known), the central planning is more effective, and we have here a limitation of effectiveness for autonomous control methods. In Sect. 4.2.2 we have shown, that the optimal controls are the piecewise constant functions, consequently, the computation of the optimal control, if values of inputs ui , i D 1; : : : ; n are available on time span Œ0; T , is resolved to the computation of the points, where the controls are discontinuous, what essentially decreases the computation time, needed for a finding of an optimal control. This shows, that for the systems, for which an extensive information is available, the central control has advantages over autonomous control. As long-term strategy (when information about ui increases with time), the model predictive control (see, e.g., [2]) can be used. But if the information is not available, the computation of optimal controls is not possible, and we have to exploit autonomous control strategies. Of course, at the same time different types of limitations may become important. For example, in the large-scale networks with high availability of information about subsystems we have, that the central planning may become impossible because of the large size of a network (essential limitation of central control), but also an autonomous control has its limitations of effectiveness, because the possibility to use the exact information from all nodes is disregarded. In this case different mixed strategies may be applied to find the most effective combination of central planning and autonomous control. For example, the results of Sect. 4.2.2 can be used to design the autonomous control method.
4.5 Conclusions We have considered two methods of control of production networks: autonomous and central control. We have shown, that if information about the flows of goods to the system from the outside is known, then the central control has certain advantages in comparison with autonomous control. In particular, in the Sect. 4.2.2 we have shown, that (in some sense) optimal distributional coefficients have to be piecewise constant functions. From the other side, if the information about the flows of goods is not available, the central methods loss their effectiveness, and the autonomous control come to a foreground. We have also outlined the recent research of stability of logistic networks, based on the notion of input-to-state stability, and discussed perspectives of future investigations in this field. In particular, the theorems were provided, that give sufficient conditions for stability of the network (4.2) in terms of distributional coefficients, that characterize the material flows between the nodes of a network. The limitations of both central and autonomous control have been discussed. We argue, that usually in real systems the mixed strategy should be used, depending on the properties of the systems, some of which have been considered in this paper.
42
S. Dashkovskiy et al.
References 1. Alekseev VM, Tikhomirov VM, Fomin SV (1987) Optimal control. Contemporary Soviet Mathematics, Consultants Bureau, (trans: from the Russian by Volosov VM) New York 2. Allgöwer F, Zheng A (eds) (2000) Nonlinear model predictive control, Progress in Systems and Control Theory, vol 26. Birkhäuser, Basel, papers from the workshop held in Ascona, June 2–6, 1998 3. Alvarez E (2007) Multi-plant production scheduling in SMEs. Robot Comput Integrated Manuf 23(6):608–613 4. Armbruster D, de Beer C, Freitag M, Jagalski T, Ringhofer C (2006) Autonomous control of production networks using a pheromone approach. Physica A 363(1):104–114 5. Armbruster D, Degond P, Ringhofer C (2006) A model for the dynamics of large queuing networks and supply chains. SIAM J Appl Math 66(3):896–920 6. Dashkovskiy S, Kosmykov M (2009) Stability of networks of hybrid ISS systems. In: Proc. of CDC/CCC 2009, The Joint 48th IEEE Conference on Decision and Control and 28th Chinese Control Conference, Shanghai, P.R. China, pp 3870–3875 7. Dashkovskiy S, Mironchenko A (2010) On the uniform input-to-state stability of reactiondiffusion systems. To appear in Proceedings of the 49th IEEE Conference on Decision and Control, Atlanta, Georgia, USA, Dec. 15–17, 2010 8. Dashkovskiy S, Naujok L (2010) Lyapunov-Razumikhin and Lyapunov-Krasovskii theorems for interconnected ISS time-delay systems. In: Proceedings of the 19th MTNS 2010, Budapest, Hungary, July 5–9, pp 1179–1184 9. Dashkovskiy S, Rüffer BS, Wirth FR (2007) An ISS small gain theorem for general networks. Math Control Signals Syst 19(2):93–122 10. Dashkovskiy S, Görges M, Kosmykov M, Mironchenko A, Naujok L (2010) Modelling and stability analysis of autonomous controlled production networks. To appear in Logistics Research 11. Dashkovskiy S, Karimi HR, Kosmykov M, Mironchenko A, Naujok L (2010) Application of the LISS Lyapunov-Krasovskii small-gain theorem to autonomously controlled production networks with time-delays. In: Proceedings of SysTol, Nice, France, Oct. 06–08, 2010, pp 765–770 12. Dashkovskiy S, Kosmykov M, Naujok L (2010) Stability of interconnected impulsive systems with and without time-delays using Lyapunov methods. Http://arxiv.org/abs/1011.2865 13. Dashkovskiy SN, Rüffer BS (2010) Local ISS of large-scale interconnections and estimates for stability regions. Syst Contr Lett 59(3-4):241–247 14. Dashkovskiy SN, Rüffer BS, Wirth FR (2010) Small gain theorems for large scale systems and construction of ISS Lyapunov functions. SIAM J Contr Optim 48(6):4089–4118 15. Kim JH, Duffie NA (2004) Backlog control for a closed loop PPC system. Ann CIRP 53(1):357–360 16. Sauer J (2006) Modeling and solving multi-site scheduling problems. In: van Wezel W, Jorna R, Meystel A (eds) Planning in intelligent systems: aspects, motivations and method, Wiley, Hoboken, NJ, pp 281–299 17. Scholz-Reiter B, Freitag M, de Beer C, Jagalski T (2005) Modelling dynamics of autonomous logistic processes: discrete-event versus continuous approaches. CIRP Ann Manuf Technol 54(1):413–416 18. Scholz-Reiter B, de Beer C, Freitag M, Jagalski T (2007) Analysing the dynamics caused by autonomously controlled logistic objects. In: ElMaraghy HA, Zaeh MF (eds) Proceedings of 2nd International Conference on Changeable, Agile, Reconfigurable and Virtual Production (CARV 2007), Windsor, pp 273–280 19. Scholz-Reiter B, de Beer C, Jagalski T (2007) Selbststeuerung logistischer Prozesse in Produktionsnetzen. Ind Manag 23(1):19–23 20. Scholz-Reiter B, Jagalski T, Bendul J (2008) Autonomous control of a shop floor based on bee’s foraging behaviour. In: Haasis H-D, Kreowski H-J, Scholz-Reiter B (eds) Proceedings of the 1st International Conference on Dynamics in logistics. LDIC 2007, Springer, Berlin, Heidelberg, pp 415–423
4
Autonomous and Central Control of Production Networks
43
21. Scholz-Reiter B, Görges M, Jagalski T, Mehrsai A (2009) Modelling and analysis of autonomously controlled production networks. In: Proceedings of the 13th IFAC Symposium on Information Control Problems in Manufacturing (INCOM 09). Moscow, Russia, pp 850–855 22. Scholz-Reiter B, Görges M, Philipp T (2009) Autonomously controlled production systems Influence of autonomous control level on logistic performance. CIRP Ann Manuf Technol 58(1):395–398 23. Scholz-Reiter B, Dashkovskiy S, Görges M, Naujok L (2010) Stability analysis of autonomously controlled production networks. To appear in Int J Prod Res 24. Scholz-Reiter B, Görges M, Jagalski T, Naujok L (2010) Modelling and analysis of an autonomous control method based on bacterial chemotaxis. In: Proceedings of the 43rd CIRP International Conference on Manufacturing Systems (ICMS 2010). Neuer Wissenschaftlicher, Wien, Austria, pp 699–706 25. Scholz-Reiter B, Lensing T, Görges M, Dickmann L (2010) Classification of dynamical patterns in autonomously controlled logistic simulations using echo state networks. Harbor, Maritime & Multimodal Logistics Modelling and Simulation (in press) 26. Scholz-Reiter B, Rekersbrink H, Görges M (2010) Dynamic flexible flow shop problems scheduling heuristics vs. autonomous control. CIRP Ann Manuf Technol 59(1):465–468 27. Tröltzsch F (2010) Optimal control of partial differential equations, Graduate Studies in Mathematics, vol 112. American Mathematical Society, Providence, RI, theory, methods and applications, (trans: from German by Jürgen Sprekels) from the 2005 28. Wiendahl HP, Lutz S (2002) Production in networks. Ann CIRP Manuf Technol 51(2):1–14 29. Windt K (2006) Selbststeuerung intelligenter objekte in der logistik. In: Vec M, Hütt M, Freund A (eds) Selbstorganisation - Ein Denksystem für Natur und Gesellschaft, Köln: Böhlau Verlag 30. Windt K, Böse F, Philipp T (2005) Criteria and application of autonomous cooperating logistic processes. In: Gao JX, Baxter DI, Sackett PJ (eds) Proceedings of the 3rd International Conference on Manufacturing Research. Advances in Manufacturing Technology and Management
Chapter 5
Approaching the Application Borders of Network Capacity Control in Road Haulage Jörn Schönberger and Herbert Kopfer
5.1 Introduction Freight forwarding companies as well as freight carriers work together to fulfill the transport demand of customers in road-based freight transportation (road haulage). A freight forwarder receives the customers’ demand and organizes a transportation chain through a given transportation network. Often, a forwarding company has no own transport equipment to realize the complete transport chain from the customer-specified pickup location to the desired delivery point. In this case, the freight forwarder organizes parts or the overall transport chain by hiring third parties that fulfill the physical movement of goods. The third parties are involved into the transshipment facilities and the transport operations. The latter operations are fulfilled by freight carriers who are engaged by the freight forwarder to move customers’ goods with their equipment (trucks, trailers, vans. . . ). Here, the demand is expressed as a request that can only be fulfilled in complete or not at all. Within this article, we analyze the interactions of freight forwarders with carriers from the view point of the carrier during the hiring negotiations. In particular, we investigate how the carrier’s decision making with respect to the request acceptance can be supported by defining and solving adequate mathematical optimization models. Capacity control subsumes efforts to scan an incoming stream of consecutively specified requests for the utilization of scarce resources with the goal to filter the most profitable requests out from the proposed stream [10]. Basic components of capacity control systems are the control of utilized resource capacities and the prediction of expected future profits (revenues). Both components are endangered by a minor data quality. The forecast of the number of future requests might be wrong so that the required capacity is incorrectly calculated or carrier dispatchers grant unexpected ad-hoc discounts to hesitating customers in order to increase profit margins. We suppose that the exactness of the incorporated demand forecast as well as the J. Schönberger (B) and H. Kopfer Chair of Logistics, University of Bremen, Bremen, Germany e-mail:
[email protected] M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_5,
45
46
J. Schönberger and H. Kopfer
frequency of granted discounts strongly correlates with the achieved overall profits. In this context, we state the following research hypothesis, which we want to verify in the here reported research: If each incoming request is inspected individually with respect to the expected profit then a better compensation of forecast errors (concerning the expected number of upcoming requests and the expected revenues) is achieved compared to the case in which a specified number of requests is accepted without further individual inspection on arrival. We start our report with the description and analysis of the decision situation (Sect. 5.2). Afterwards, we propose capacity control approaches for the investigated dynamic decision problem (Sect. 5.3). A simulation study is reported, in which the proposed capacity control approaches are incorporated, the observed results are discussed and an answer to the motivating research hypothesis is derived (Sect. 5.4).
5.2 Decision Situation A compilation of literature related to our investigation subject initializes our research (Sect. 5.2.1). Then, we give an informal description of the analyzed decision problem (Sect. 5.2.2) and a decision model (Sect. 5.2.3). Finally, we describe the construction of artificial test cases (Sect. 5.2.4) that will be used to evaluate the capacity control approaches proposed later on in this report.
5.2.1 Literature There are a lot of scientific publications dealing with operational transport process planning for road-based freight transportation. These contributions are classified into three subproblems. The most often referred subproblem class in operational freight transportation comprises the vehicle routing problem and its variants [7], in which least cost routes and schedules are compiled for the available vehicles. The externalization of subsets of a request portfolio (subcontracting) is also investigated [11]. Often, these two subproblems of operations transport process planning are considered as an integrated problem [12, 14]. The third referred subproblem is the request acceptance problem. Here, the dispatcher has to select a subset of requests from a given request portfolio. Chosen requests are compiled into routes or externalized and fulfilled but the remaining requests are rejected. Static and dynamic request acceptance situations are distinguished. In a static situation, the overall request portfolio is known at the time when the acceptance decision is made. Typically, variants (extensions) of the vehicle routing problem like the selective vehicle routing problem [1, 21], the profitable salesman problem [4] or pickup and delivery selection problems [18] are investigated. If the portfolio is not known completely at the acceptance decision time, the acceptance
5
Approaching the Application Borders of Network Capacity Control
47
problem is dynamic. Whenever additional requests are released, it has to be decided again, which subset of the additional requests is accepted and which requests are rejected. Dynamic acceptance problems belong to the class of capacity control problems [20] which are an important decision problem class in operational revenue management [3]. The basic challenge in capacity control is to filter an incoming stream of requests so that a scarce resource is used in the most profitable way by the selected requests. Network capacity control considers several linked resources simultaneously [9]. The major challenge in capacity control is that a once made acceptance decision cannot be revised later on. However, if flexible products are offered, it is possible to alter the request fulfillment process after a request has been definitively accepted [6, 13]. In road based freight transportation, a product is flexible, if there are two or more routes or services to fulfill a request. This dispatching option enables a network service provider (carrier) to increase its profitability [15]. Capacity control problems are typically represented as dynamic programming (DP) models. However, the tree structure required to store and evaluate all future system states in the DP is too large to be efficiently handled. Thus, so called deterministic linear programs (DLP) are used as decision models [10]. In a DLP, the evaluation of the future states is simplified. Only real-valued expected values (e.g. the expected number of upcoming requests or the expected sum of expected revenues collectible from the upcoming requests) are considered to represent future system states. Capacity control strategies represent rules that describe under which circumstances a request is accepted [16]. Two forms of strategies are discussed in the scientific literature. At first, quotes are determined for products specifying the maximally allowed number of accepted requests [17]. A proposed request is accepted without further individual inspection as long as the quota is not exceeded. At second, in a bid-price approach [9], each request is individually checked for profitability at its arrival time. However, the determination of the required margin costs (“bidprice”) for requests is a very challenging problem. Several concepts are proposed [1, 9] to approximate these costs. Within this report we develop a network capacity control system and evaluate a quota- as well as a bid-price approach as decision support approach for a road freight carrier.
5.2.2 Dynamic Decision Problem A road freight carrier is confronted with an incoming stream of consecutively arriving request proposals originated by different freight forwarders searching for transport resources. Each request expresses the indivisible demand for the movement of goods of a certain quantity through the network N WD .V; A/ formed by the set of pickup and delivery points V and connections between these nodes collected in the set A.
48
J. Schönberger and H. Kopfer
The definition of N is based on planned transportation services necessary to fulfill demand specified in longer term contracts. A regular visit of the nodes in V by a truck of the carrier is necessary for loading and/or unloading goods according to the previously mentioned contracts. To enable the fulfillment of the long term contracts the carrier operates regular transport services. A service s is an ordered sequence s D .s1 ; : : : ; sn.s/ / of n.s/ nodes taken from V . It is fulfilled by a truck that offers a given capacity CAPs . This truck visits the nodes s1 ; : : : ; sn.s/ in the defined sequence. All operated services are collected in the set S . The arcs a.s; 1/ D .s1 ; s2 /; a.s; 2/ WD .s2 ; s3 /; : : : ; a.s; n.s/ 1/ WD .sn.s/1 ; sn.s/ / form the arc set A. Each arc .si ; sj / has been indexed with the donating service s in order to have several copies of .si ; sj /;which are offered by different services, in the set A. An arc a 2 A is referred to as resource and its capacity C.a/ is initialized by C.a/ WD CAPs . For a given resource a 2 As.a/ refers to the service that provides a. Products are derived from the specified services running in N . A product p is an ordered pair of two nodes p WD .u; v/ 2 V V so that there is at least one service s 2 S , which is able to pick up some goods at u and moves them without transshipment to v where the goods are unloaded. All derived products are collected in the set P . The binary indicator SPR(s,p,a) is set to 1 if and only if service s 2 S offers product p 2 P and, doing so, uses the resource a 2 A (otherwise SPR(s,p,a): D 0). A request r expresses the demand for a specific product p WD P .r/ 2 P . The carrier expects a profit REV(P(r)) from the execution of request r. At time t k .k D 0; 1; : : : ; Kmax/ several spot-market requests are proposed to the carrier by one or more forwarding companies. These requests are proposed in addition to the already known regular demand described in the long-term contracts. The carrier is free to decide if it wants to take the fulfillment responsibility for a spot-market request. However, the carrier has to decide immediately about the acceptance of the recently proposed spot market requests (especially, it cannot wait until further requests become known). We assume that the acceptance of a spotmarket request does not lead to considerable additional accountable route execution costs. Thus, the carriers total profit is lifted by the collected revenue REV(P(r)) associated with the spot-market request r. The carriers’ goal is now to filter the most profitable requests from the stream of consecutively arriving requests considering the already made irreversible decisions as well as the scarce capacity of the available resources. Therefore it decides about the acceptance/rejection of the currently appeared requests guided by a capacity control strategy. The carrier derives its decisions from calculated bid-prices of the available resources or from previously fixed product- and service-specific quotes. Thus, the carriers’ decision problem can be interpreted as a dynamic bid-price and/or quota determination problem. At the decision time t k , the carrier knows the achievable revenues REV(P(r)). associated with a request r, the number Y k .P .r// of requests for product P .r/ already accepted not later than t k1 , the expected demand-to-come DTCk .P .r// for product P .r/ in the remaining booking phase Œt k ; t Kmax as well as the remaining capacity of the available resources. It must be found out if r is accepted or if the potentially scarce capacity of the resources will be assigned to later arriving
5
Approaching the Application Borders of Network Capacity Control
49
requests, which are more profitable but whose appearance is unsure (which results in the rejection of the currently waiting request r). Although the carrier is unable to revise the acceptance decision of a once accepted or rejected request it may revise the assignment of an already accepted request r from a service s 0 to another service s if the product P .r / is flexible, e.g. if two or even more services offer product P .r/ and provide free capacity.
5.2.3 Model-Based Determination of Quotes and Bid-Prices We represent the previously outlined dynamic decision problem as an online decision model M D .M1 ; M2 ; : : : ; M Kmax / [19]. Whenever additional requests arrive at time t k we adjust the stored bid-prices BP(a) of all resources a 2 A and the service-specific quotes Q.p; s/.p 2 P; s 2 S / for the products to the current workload and to the expected future demand. In order to define the decision model instance M k at time t k we introduce the two decision variable families ypk .p 2 P / and zkp;s .p 2 P , s 2 S ) to code the revenue maximal quotes for each product p (the y-variables) and the service-specific quotes (the z-variables). X
REV .p/ ypk ! max
(5.1)
SPR.s.a/; p; a/ zkp;s.a/ C.a/ 8a 2 A
(5.2)
zkp;s D ypk
(5.3)
p2P
X
p2P
X
8p 2 P
s2S
ypk Y k .p/
8p 2 P
ypk Y k .p/ C DT C k .p/ ypk ; zkp;s 0
(5.4) 8p 2 P
8p 2 P; s 2 S
(5.5) (5.6)
The total sum of revenues is maximized (5.1) by finding adequate quotes ypk for each product p 2 P . Gallego et al. (2004) propose the linear constraints (5.2)–(5.6) to encode the collections of feasible quotes ypk as well as zkp;s (cf. [5]). Constraint (5.2) ensures that the capacity of the considered resources is not exceeded. All accepted requests must be distributed among the available services that can fulfill the considered requests (5.3) and it is guaranteed for each already accepted request that enough capacity is reserved for each product p so that there is at least one service for fulfilling the already accepted requests associated with product p (5.4). We do not reserve capacity for more requests than expected to appear during the remaining booking phase Œt k ; t Kmax (5.5). Klein (2007) explains why it is not necessary to enforce the integer-property of the quota decision variables in the model [9].
50
J. Schönberger and H. Kopfer
The model (5.1)–(5.6) is a DLP, it is denoted as M k . Information about the uncertainty and value of the upcoming spot-market requests are compressed in the expectation values DT C k .p/.p 2 P /. We (re-)solve the instance M k of this program at every time point t k .k D 1; : : : ; K max / to update the so far used quotes Q.p; s/ with respect to the meanwhile arrived requests and to the still available resources. After the solution of (5.1)–(5.6) has been determined, we update the quotes Q.p; s/ by setting Q.p; s/ WD zkp;s for all products p 2 P and all services s 2 S . If we use a simplex algorithm to solve the DLP (5.1)–(5.6) we are able to fetch shadow prices of all resources a 2 A from the final simplex tableau. For a given resource a the shadow price equals the entry in the final simplex tableau that is found in the objective function row in the column of the slack variable belonging to the restriction of a’s capacity in the constraint family (5.2). The shadow price spk .a/ of a resource a represents opportunity costs for the utilization of one unit of this resource. Hence, the shadow price of resource a represents the least necessary revenue that must be earned if one unit of this capacity is used in order to increase the overall sum of earned revenues. Consequently, the shadow price of a resource represents a reasonable bid price for this resource at time t k . We update BP(a) at time t k by setting BP(a): D spk .a/ for all a 2 A. In conclusion, the solving of the DLP (5.1)–(5.6) leads to the update of servicespecific product quotes as well as to the update of resource-specific bid-prices. The updated values of Q.p; s/ and of BP(a) have been calculated under consideration of all knowledge acquired until time t k and under consideration of the expected demand for the remaining booking period from t k to t Kmax .
5.2.4 Test Cases We want to simulate the carrier’s decision making during the scaled booking phase Œ0I 1 for services operating on a particular day in a given network N . To define artificial test cases, we first construct a network N: D (V,A). The vertice set V is formed by 18 major cities in Germany [2]. The cities are labeled by i D 1; : : : ; 18. There are five offered services S WD fs1 ; : : : ; s5 g specified for the considered execution day. The defined services are s1 WD .2; 9; 11; 15; 6; 10; 8; 4; 16; 5; 12/, s2 WD .2; 1; 13; 14; 12; 3/, s3 WD .1; 11; 15; 6; 5; 17; 7/, s4 WD .1; 11; 18; 5; 12; 3/ and s5 WD .10; 6; 18; 15; 14; 13; 12; 7/. The offered services in the set S determine the available and bookable resources. In the current setting, the offered services provide 32 resources. The capacity C.a/ of each resource a 2 A is set to 25 capacity units. Overall, the product set P comprises 106 products that are offered by the five services (Table 5.1). Among them, we find 20 flexible products that are indicated by F as well as 86 simple products indicated by S.
5
Approaching the Application Borders of Network Capacity Control
51
Table 5.1 Products induced by the five services Delivery location
Pickup location
1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
S
2
3 F S
4 S
5 F S
6 S S
7 S
8
9
10
S
S
S
11 F S
S S
S F S S
F
F
S
S
S S S S
S S S F
S S F
S S S
S
S
S
S F F
S S
F S S
F
S S S S S S S S
S
S
S
12 F F
S
S S F F
13 S S
15 S S
16
17 S
18 S
S
S S
S
S S S S
S
S S
S
S
S S
S
S
S S
S F
F F S S F
14 S S
S
S
S
As commonly done in the context of revenue management investigations, we assume that incoming booking requests arrive according to a non-homogeneous Poisson process (cf. [8]). For each product p 2 P eleven potential replanning time points t k D 0; 0:1; 0:2; : : : ; 1 during the booking phase are assumed. The carrier checks for additionally arrived requests only at these time points and it decides about the acceptance or rejection of the proposed spot-market requests. At time t k the number of incoming requests for product p is p .t k / (demand intensity). Up to t WD tp;st art the demand intensity equals 0. Next, p .t/ climbs up proportionally until the maximal demand intensity p;max is reached at time point tpmax P . Afterwards, the intensity declines linearly until it reaches 0 at time t WD tp;finish (cf. Fig. 5.1). The resulting triangle-shaped intensity functions p .t/ are called booking curves of the products p 2 P [9]. If we determine the total number of requests to be released during the booking phase Œ0I 1, the time points tp;start , tp;finish and tp;max then the intensity function p is well defined. For all products, we set tp;start WD 0:1, tp;max D 0:5 and tp;finish D 1:0 D t Kmax . To define a demand pattern for product p, we consecutively draw the number bp .t/c of arriving spot-market requests for each time point t k D 0; 0:1; : : : ; 1 according to the booking curve p . Then, we generate bp .t/c requests of product p and set their arrival times to t k . After the requests have been generated for all products p 2 P , the generated request lists are merged and the merged list is sorted
52
J. Schönberger and H. Kopfer
Fig. 5.1 Intensity function p(t) of product p
by increasing arrival times. The capacity consumption of a request is set to one capacity unit for each used resource. A demand forecast .˛; ˇ/ is added to a demand pattern and the triple (; ˛; ˇ) then forms a scenario. The first parameter ˛ determines the correctness of the forecast of the number of incoming requests. If ˛ D 1 then the used forecast is exact, but if ˛ > 1 then an overestimation of the number of requests is made (e.g. ˛ D 1:05 means an overestimation of 5%). An ˛-value below 1 represents an underestimation of the number of incoming requests. The second parameter ˇ determines the congruence of the actually achieved revenue associated with an accepted request r and of the revenue expected to be gained from r during the derivation of bidprices and quotes. An ad-hoc discount of 25% applies to an incoming request with probability ˇ. The formal specification of the subject of investigation enables the refinement of the initially stated research question: We want to demonstrate that in network capacity control a bid-price capacity control strategy exhibits a better performance than a quota-based capacity control strategy if ˛ >> 1 or ˛ > 0 (unreliable revenue forecast) which means that a bid-price capacity control strategy is able to collect more revenues from the same incoming request proposal stream than a quota-based capacity control strategy is able to do.
5.3 Capacity Control System We start with an outline of the proposed capacity control system (Sect. 5.3.1). Next, a decision model for the tentative assignment of already accepted requests is proposed in Sect. 5.3.2. The definition of control policies completes the description of the capacity control system (Sect. 5.3.3).
5
Approaching the Application Borders of Network Capacity Control
53
Fig. 5.2 Pseudo-code of the capacity-control system
5.3.1 Outline of the System The procedure capacity_control() whose pseudo-code is shown in Fig. 5.2 is used to process the online model (M 1 ; : : : ; M Kmax /. At first, the network N is determined, the decision times are fixed and the capacity control strategy is specified (a). Next, the iteration counter k is initialized (b) and the set of already accepted requests is inaugurated (c). For every decision time t k the loop (d)–(o) is executed. The current time is fetched first (e). Then a tentative resource allocation is made for the already accepted requests in order to determine the still available capacities of the resources in A (f). It is continued with the definition of the current instance of the parameter update model (5.1)–(5.6) in line (g). This model is solved (h) and the quotes as well as resource bid prices are updated. Now, the currently waiting spot-market requests are collected in the set REQk (i). If a quota-based capacity control strategy is used then the waiting requests are processed calling the decision making procedure process_by_quota()(j) but if a bid-price-based capacity control strategy is incorporated then the processing of the fetched requests is done by calling the procedure process_by_bp()(k). After all requests have been processed the acceptance decisions are sent back to the customers (l). The set ACC of already accepted requests is updated (m) and the iteration counter is increased by 1 (n).
5.3.2 Resource (Re-)Allocation When the acceptance decision is made for a request then it is ensured that enough capacity is available and a tentative resource allocation is made. Since the demand
54
J. Schönberger and H. Kopfer
forecast and/or the expected revenue differ from the actually appearing demand it is necessary to evaluate and re-work the assignment regularly in order to ensure an optimal resource utilization. Therefore, the procedure make_tentative_assignment(ACC,N) is called. Its general function is described in this section. At time t k we have to assign each accepted request r 2 AC C to an appropriate service s 2 S operating in the considered network N . The limited capacity of each service must be considered. For coding the necessary allocation decisions we introk duce the family xrsk (r 2 AC C; s 2 S / of binary decision variables. We define xrs k to be 1 if and only if request r is assigned to service s at time t . In a preprocessing step, we determine the value of the indicators .r; s/. If service s offers the product of request r we set .r; s/ to 1 otherwise we set .r; s/ to 0. It is necessary to assign each request to exactly one service. In order to be able to allocate resources belonging to service s for request r it is necessary to limit the entirety of selectable services. Only those services offering product P .r/ become potential candidate services for request r. The capacity of the transport resources a 2 A are limited and only CAP(a) capacity units can be moved at the same time by a resource. X
xrsk D 1 8r 2 AC C
(5.7)
s2S
xrsk .r; s/ 8r 2 AC C; s 2 S X X SPR.S.a/; P .r/; a/ xrsk C.a/ 8a 2 A
(5.8) (5.9)
s2S r2AC C
xrsk 2 f0; 1g
8r 2 AC C; s 2 S
(5.10)
Every assignment scheme that respects the constraints (5.7)–(5.10) is a feasible one at the replanning time t k . A partition of the current portfolio of accepted requests is achieved by incorporating constraint (5.7). The consideration of constraint (5.8) ensures that request r is executable by service s. Constraint (5.9) must be satisfied in order to respect the limited capacities of the resources. Each accepted request must be assigned to a service in complete or not at all. X X
BP .r; s/ xrsk ! min
(5.11)
r2AC C s2S
Among the available allocations represented by the feasible assignment schemes we are looking for a most beneficiary one. In our investigation, we evaluate an allocation by its profit. Therefore, we determine the assignment specific costs BP(r, s) that account if resources of service s are allocated for the fulfillment of request r. We approximate BP(r, s) by summing up the current bid-prices of resources belonging to service s and being deployed in the fulfillment of r. The overall costs for an assignment scheme are achieved by summing up the BP(r, s)-values for each selected assignment. We select the assignment scheme that minimizes this sum
5
Approaching the Application Borders of Network Capacity Control
55
(5.11) while fulfilling (5.7)–(5.10). This resulting optimization problem (5.7)–(5.11) is a multi-commodity network flow problem.
5.3.3 Control Policies This section outlines the procedure that processes the additionally arrived requests and that determines the required acceptance decisions.
5.3.3.1 Quota-Based Control (QUOTA) The solving of the DLP (5.1)–(5.6) determines service- and product-specific contingents Q.p; s/ (quotes). In order to earn the maximal possible revenue from the stream of requests from the spot-market an additionally arrived request for product p using service s is accepted as long as the contingent Q.p; s/ is not completely exhausted. At first, the procedure process_by_quota() selects an arbitrary order of the requests contained in REQk . These requests are then processed one after another. In order to decide whether request r is accepted, the procedure determines the product P .r/ of request r first. Next, it checks if there are still service(s) available that can fulfill r. For that purpose, it tests consecutively for all available services s, if Q.P .r/; s/ > 0. Let s be the first service whose quota is still available. If such a service does not exist, r is rejected. Otherwise, r is accepted and the quota is updated by Q.P .r/; s / WD Q.P .r/; s / 1. Request r is tentatively assigned to service s . 5.3.3.2 Bid-Price Control (BP) Also the procedure process_by_bp() first serializes the waiting requests and then processes the requests one after another. According to commonly used acceptance policies for flexible products [15], the carrier accepts r if and only if there is at least one service s available that can be used to fulfill r and that has a bid-price BP(r, s / below the revenues REV(P(r)) associated with the fulfillment of r. Hence, r is accepted if and only if BP .r; s / REV .P .r// and if sufficiently capacity is available for service s . Request r is then tentatively assigned to the service s that leads to the highest positive contribution margin REV .P .r// BP .r; s /.
5.4 Computational Experiments This section reports about computational experiments in which the capacity control system developed in Sect. 5.3 is evaluated. The booking processes for the demand scenarios introduced in Sect. 5.2.4 are simulated. Metrics used to represent the
56
J. Schönberger and H. Kopfer
simulation results are introduced in Sect. 5.4.1 and the observed results are presented and discussed in Sect. 5.4.2.
5.4.1 Evaluation Metrics We execute a comprehensive evaluation of the two capacity control policies QUOTA and BP introduced in Sect. 5.3.3. To facilitate such an evaluation, we define different performance indicators in order to represent the policy behavior during the normalized booking period Œ0I 1 with the acceptance decision time points 0; 0:1; : : : ; 1. From time t D 1 on, the values of the performance indicators do not vary anymore since the booking phase has been closed. Thus, the discussion of the performance indicator values at time t D 1 enables an ex-post description of the control policies’ performance (offline analysis). We fetch the sum of revenues rev1 .i; ; ˛; ˇ/ gained up to time 1 for the i -th realization (draw) of a demand pattern and with the applied control policy 2 fQUOTA; BP g and the forecast of quality (˛; ˇ). For each simulation experiment configuration (; ˛; ˇ) we calculate the average rev1 .; ˛; ˇ/ over all demand pattern draws. Let revmax .˛; ˇ/ be the P : : : ; 1g; 2 fQUOTA; BP gg. The maximum of the set frevt .; ˛; ˇ/jt 2 f0; 01; normalized sum of gained revenues r1 .; ˛; ˇ/ for the configuration .; ˛; ˇ/ is 1 .;˛;ˇ / then defined as r1 .; ˛; ˇ/ WD rev . These values vary between 0 and 1. revmax .˛;ˇ /
5.4.2 Results Both policies exhibit a similar performance with regard to the gained revenues. The absolute sum of gained revenues achieved for the BP-policy is slightly below the best result for the QUOTA-policy and the highest observed sum of gained revenues is observed for the QUOTA-policy if the forecast is nearly perfect (˛ 0:9; ˇ 0). The normalized revenue sum rt .; ˛; ˇ/ uncovers revenue shifts if the accuracy of demand forecasts varies and the two isoline-plots in Fig. 5.3 compile the values of r1 .; ˛; ˇ/. We observe the highest gained revenue sum if no ad hoc discounts are granted to customers (ˇ D 0) and if we have a slightly too low forecast of the number of requests (˛ D 0:9). For both control policies we detect a significant decrease of the gained revenues if we allow ad hoc discounts (increase of ˇ) and if the forecast of the number of incoming requests is quite too high (˛ > 1:0) or quite too low (˛ < 0:9). In summary, we have found out that the revenue loss grows up if the accuracy of the forecast (of the request number and correct revenue) declines. A wrong expectation of the expected revenue takes place in the scenarios with ˛ 0:7 and ˇ 0:5 (WR D Wrong Revenue). In the WR-case, both control
5
Approaching the Application Borders of Network Capacity Control
57
Fig. 5.3 Remaining percentage of revenues r1 .QUOTA; ˛; ˇ/ (left) and r1 .BP; ˛; ˇ/ (right)
policies show a similar behavior. The revenue decrease r1 is between less than 20% up to 40%. If the quality of the forecast falls into the High-Quality-(HQ)-area (˛ 0:7 and ˇ < 0:5) then we observe a decrease of the achieved revenues down to 80% in the worst case. However, if the accuracy ˛ of the forecast of the request number is between 0.8 and 1 then a relative remaining revenue level of 90% is observed up to a frequency of discounted requests of 30% (QUOTA) or 40% (BP). In the WN-area (scenarios with a wrong prediction of the expected number of incoming requests, ˛ < 0:7 and ˇ < 0:5) the performance difference between BP and QUOTA becomes obvious. The bid-price approach is able to keep r1 .BP; ˛; ˇ/ above 70% but at the same time, we observe a decrease of r1 .QUOTA; ˛; ˇ/ down to 30% for the QUOTA approach. If the request number forecast accuracy ˛ falls below 50% then the sum of gained revenues collapses in the QUOTA-experiments. Finally, different revenue sums are observed in the LQ-area (Low-Quality-area, ˛ < 0:7 and ˇ 0:5). Here, the application of the BP-policy bounds the decrease of the revenue sum by 40% but the QUOTA-policy is unable to maintain this revenue sum. Finally, a revenue decrease of 70% is observed (down to r1 .QUOTA; 0:2; 1/ D 0:31 Remarkably, in most of the experiments in the LQ-area revenues of 70% and even 80% of the best observed revenue sum are collected by the BP-policy. In summary, the QUOTA-approach is more sensible against forecast errors than the BP-approach is with respect to the gained revenues during a booking phase. Vice-versa, we have proven that the BP control policy is able to compensate prognosis faults to a larger extend than the QUOTA policy is able to do. We conclude from this observation that the BP approach is more suitable to serve as a basic tool in the forwarder/carrier-negotiations if reliable demand forecasts are not available. In conclusion, these observations verify the initially stated (and in Sect. 5.2 refined) research hypothesis with respect to the gained revenues: The BP-police
58
J. Schönberger and H. Kopfer
outperforms the QUOTA-policy and collects more revenues if the demand forecasts are erroneous.
5.5 Conclusions Within the reported research we have proposed a capacity control system for supporting the sales and capacity management of scheduled services in road-based freight transportation. A quota-based capacity control strategy and a bid-pricefounded capacity control strategy have been proposed. Within comprehensive computational simulation experiments their general applicability has been proven. We were able to proof our research hypothesis that it is beneficial to inspect each incoming requests individually instead of determining contingents for products if the available demand forecasts are erroneous. Future research efforts will address the impacts of wrong demand volume forecasts. Furthermore, we will investigate how short-term variations of the services in the network influence the performance of the capacity control strategies.
References 1. Aras N, Aksen D, Tekin M (2010) Selective multi-depot vehicle routing problem with pricing. Transport Res C Emerg Technol to be published. doi: 10.1016/j.trc. 2010.08.003 (in press) 2. Becker M, Wenning BL, Görg C, Gehrke JD, Lorenz M, Herzog O (2006) Agent-based and discrete event simulation of autonomous logistic process. In: Borutzky W, Orsoni A, Zobel R (eds) 20th European Conference on Modelling and Simulation, 56671 3. Corsten H, Gössinger R (2005) Kapazitätssteuerung im revenue management. Zeitschrift für Betriebswirtschaftslehre 67:31–52 4. Feillet D, Dejax P, Gendreau M (2005) Traveling salesman problems with profits. Transport Sci 39:188–205 5. Gallego G, Iyengar G, Phillips R, Dubey A (2004) Managing flexible products on a network. Technical report, Ithaca, NY, USA 6. Gallego G, Phillips R (2004) Revenue management of flexible products. Manuf Service Oper Manag 6:321–337 7. Golden B, Raghavan S, Wasil E (2008) The vehicle routing problem. Springer 8. Kimms A, Müller-Bungart M (2007) Simulation of stochastic demand data streams for network revenue management problems. OR Spectrum 29:5–20 9. Klein R (2007) Network capacity control using self-adjusting bid-prices. OR Spectrum 29:39–60 10. Klein R, Steinhardt C (2009) Revenue management. Springer 11. Kopfer H, Wang X (2009) Combining vehicle routing with forwarding - extension of the vehicle routing problem by different types of sub-contraction. J Korean Inst Ind Eng 35:1–14 12. Krajewska M, Kopfer H (2009) Transportation planning in freight forwarding companies - tabu search algorithm for the integrated operational transportation planning problem. Eur J Oper Res 197:741–751 13. Müller-Bungart M (2007) Revenue management with flexible products. Springer 14. Pankratz G (2002) Speditionelle transportdisposition. DUV, Wiesbaden, Germany
5
Approaching the Application Borders of Network Capacity Control
59
15. Petrick A, Steinhard C, Gönsch J, Klein R (2010) Using flexible products to cope with demand uncertainty in revenue management. OR Spectrum. doi: 10.1007/s00291-009-0188-1 (in press) 16. Petrick AJG, Steinhard C, Klein R (2010) Dynamic control mechanisms for revenue management with flexible products. Comput Oper Res 2027–2039 17. Rehkopf S (2006) Revenue management-konzepte zur auftragsannahme bei kundenindividueller produktion. DUV, Wiesbaden, Germany 18. Schönberger J (2005) Operational freight carrier planning. Springer, Berlin 19. Schönberger J, Kopfer H (2009) Online decision making and automatic decision model adaptation. Comput Oper Res 36:1740–1750 20. Talluri K, van Ryzin GJ (2005) The theory and practice of revenue management. Springer, Berlin 21. Valle C, da Cunha AS, Mateus G, Martinez L (2009) Exact algorithms for a selective vehicle routing problem where the longest route is minimized. Electron Notes Discrete Math 35:133–138
Chapter 6
Limit and Degree of Autonomy in Groupage Systems: Challenges, Chances and Barriers for Horizontal Cooperation in Operational Transportation Planning Heiko Wieland Kopfer, Herbert Kopfer, and Xin Wang
6.1 Motivation Due to fierce competition in the market and a growing demand for “green transportation”, forwarders in the road haulage sector are intensively trying to achieve additional means for increasing the efficiency of their fulfillment processes. Since the research on the economization of transportation has almost exhausted the potentials of traditional planning methods, progressive approaches extending the scope of usual vehicle routing are becoming more and more important for the research on transportation logistics. One possible approach for enhancing operational transportation planning is offered by collaborating within a coalition of forwarders. Such coalitions constitute a framework for horizontal cooperation and enable an extended planning procedure that is capable of incorporating request exchanges in forwarder coalitions [18]. Cooperation by exchanging transportation requests among coalition members is an option especially suitable for small and mid-sized companies that try to improve their operations, increase their market share and ensure successful future business operation. Most freight forwarding companies have to cope with a strongly fluctuating demand on the transportation market which varies considerably over time. Aside from long-term fluctuations they have to manage the daily variations of their demand. Each day a varying number of orders are received from customers on short call [19]. Additionally, some of the orders that have to be fulfilled will not suit well to the portfolio of the orders to be fulfilled during the same time horizon. The efficiency of transportation demand fulfillment can be increased tremendously by the
H.W. Kopfer (B) Institute of Machine Elements, Fastening Systems and Product Innovation, University of Siegen, Siegen, Germany e-mail:
[email protected] H. Kopfer and X. Wang Chair of Logistics, University of Bremen, Bremen, Germany
M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_6,
61
62
H.W. Kopfer et al.
possibility of sub-contracting requests to external carriers. An additional and even more powerful means, which can be applied in combination with sub-contraction, consists in harmonizing the transportation plans among partners within a coalition. Such a coalition is called a Groupage System [16]. The efficiency increase within the coalition is achieved by exchanging requests on the basis of a concerted planning process. The amount of efficiency increase reachable in Groupage Systems depends on many factors as for instance the compatibility of the partners, the organizational structures of the Groupage System, the trust shown among the partners, the mechanisms applied for request exchange, the type and amount of shared information and last not least the agreed models for cost and profit sharing. The extension of standard vehicle and routing problems introduced by the investigation of Groupage Systems transforms the usual problems to more general collaborative planning problems [16]. The horizontal cooperation between forwarders within a Groupage System provides the possibility that each partner of the Groupage System offers a part of his or her transportation requests to the coalition while each partner of the coalition can make a bid on all offered requests. A request will be forwarded to that partner who submits the best bid; i.e. to the partner who announces that he or she can fulfill the request in the cheapest way [22]. An alternative to Groupage Systems are electronic freight exchange systems. In contrast to these systems, Groupage Systems are closed systems with a limited and well-known number of participants. Sometimes they are established only among profit centers of one single company [15, 22, 29]. The partners of a Groupage System partly exchange sensitive information and purposeful adjust their plans to each other in order to reach additional synergy effects. Electronic freight exchange systems do not pursue any collaborative approach and as such only offer an electronic platform for the realization of request exchange within usual business relations. Further, Groupage Systems aim at improving the efficiency of the entire system. Thereby a common underlying objective is assumed, whereas freight exchange systems often focus only on bilateral exchanges to enhance individual situations. As such, collaborative planning considers the overall planning situation of all carriers whereas freight exchange systems focus on “selling” individual shipments. This contribution investigates the limits and the degrees of the autonomy in Groupage Systems. Some approaches for sharing different types of resources are discussed and assigned to resulting intensity-levels of cooperation. The intensity of cooperation determines the degree of autonomy of the Groupage-partners and influences the potential of positive emergence of the entire system. The following Sect. 6.2 will present an overview on collaboration in Groupage Systems. Section 6.3 will discuss prerequisites for a successful cooperation and obstacles which have to be overcome. In Sect. 6.4 the limits of collaboration will be shown. Section 6.5 is addressed to the variation of the degree of autonomy and the resulting effects on the efficiency in collaborative transportation systems. Finally, Sect. 6.6 presents some concluding remarks and an outlook.
6
Limit and Degree of Autonomy in Groupage Systems
63
6.2 Collaboration in Groupage Systems Transportation planning refers to all tasks related to the coordination of the physical movement of goods. It aims for efficient processes in terms of cost and other performance criteria such as punctuality, reliability and flexibility. According to Pfohl [26] transportation planning can be classified as a problem of operational management. Further, the tasks of inter- and intra-company transportation planning can be distinguished. Our attention is on inter-company transportation planning and on scenarios with outsourced transportation tasks. Groupage Systems offer the chance to enhance the results of transportation planning for a group of forwarders. In such systems forwarders aim at advantages by sharing requests gathered in an exchange pool and visible for all partners of the system. The increased number of disposable requests and the extended deployment portfolio resulting from cooperation bring economies of scale for each individual freight forwarder [20]. Economies of scope are created by a better capacity utilization of the available vehicles and the ability to offer the additional serving of relations which otherwise could not be handled. The general purpose of cooperation is to find equilibrium between the required and available transportation resources of the coalition partners by interchanging their customer requests [16]. Thus, in the collaborative planning process resources are directly connected, and relevant data are exchanged in order to create a common and mutually agreed plan [13]. The additional profit generated through the collaboration process is split among the coalition members according to predetermined rules. To ensure the long-term functioning of collaboration structures among independent freight carriers, positive incentives for the partners must be generated. In other words, an appropriate profit sharing scheme should guarantee a financial advantage for each partner. The features to be included in the profit sharing scheme depend on the distribution of power among freight carriers, on their level of interdependency, on their willingness to make compromises, and on the market within which they operate.
6.2.1 Literature Review A comprehensive literature review on collaboration in the transportation sector on the operational level can be found in Krajewska [20] and in Kopfer et al. [18]. The following literature review is partly extracted from these two sources. Developments in telecommunications and information technology have created many opportunities to increase cooperation among the entities operating in logistics chains. This has led to the insight that suppliers, consumers and even competitors can be potential collaboration partners in logistics Ergun et al. [7]. Vertical cooperation, involving suppliers, manufacturers, distribution centers, customers and logistics service providers has been the topic of extensive academic research [5]. This research mainly focuses on identifying potential benefits [10], critical success factors [32], and partner selection criteria [4, 28]. Formal logistics models for vertical collaboration have enabled performance improvements for an entire supply
64
H.W. Kopfer et al.
chain [30], covering the analysis of bottlenecks and the quality of customer service. Furthermore, models have been developed to predict costs and their apportionment when existing logistics chains are adapted to new products and markets [30] or when new cooperative logistics networks are designed [8,23,25]. Widely discussed issues include specific types of vertical collaboration such as cooperation models between manufacturers and retailers [11, 12, 24, 31, 33], or effective cooperation mechanisms for inventory decisions [1, 9]. In comparison, the literature on horizontal cooperation in logistics (i.e. among competitors) is still at an early stage [5]. (Only a few models of horizontal cooperation have been developed. All of these models assume an equal distribution of power and similar market positions for each of the freight carriers and focus on short-term planning. Cruijssen and Salomon [6] analyze the effect of collaboration for an entire coalition and show, using a case study, that cost savings may range from 5% to 15%. Ergun et al. [7] focus on minimizing execution costs for a coalition of freight forwarders. They assume that the goal of collaborating shippers is to identify a set of routes that can be submitted to a carrier as a bundle, instead of as individual requests, in the hope that this will result in more favorable rates. Reduced rates can be achieved when covering the routes in the bundle, and this involves little or no asset repositioning. Thus, given a set of requests to serve, it is possible to identify common tours to cover all requests and minimize asset repositioning costs. This shipper collaboration problem is considered as a constrained variant of the lane covering problem defined by Cruijssen and Salomon [6], which is similar to the cycle covering problem. Using a greedy heuristic as well as set partitioning, sets of cycles are generated which show significant cost reduction in asset repositioning. A Groupage System provides the framework for the joint operational planning of shipments. The general idea of joint operational planning is illustrated in Fig. 6.1.
Fig. 6.1 Degrees of request and shipment planning of five carriers [18]
6
Limit and Degree of Autonomy in Groupage Systems
65
The figure depicts an example of five carriers each with its transportation requests (represented by single dark dots) for the respective planning period. In this example each of the carriers is a company of the same size, however, the number of transportation requests for a planning period may vary for each carrier (e.g. Carrier 1 has three and Carrier 2 four transportation requests). Different degrees of collaborative planning are marked by different black rectangles. All transportation requests within a rectangle are potential shipments that may be exchanged with the partners. As such, the largest rectangle depicts the case of merging all requests of all carriers. In this case, collaborative transportation planning would mean that all requests are transferable shipments. This means that a planning problem including the assignment of shipments to carriers has to be solved and the resulting transportation problem for each carrier has to be solved. However, a certain reluctance of carriers to reveal details of all their customer requests is found in practice. As such, smaller numbers of requests will be selected as shipments as is exemplarily depicted by the two smaller rectangles. This variation in the number of shipments is also referred to as the modification of the degree of collaboration. It may vary from planning period to planning period. This idea of joint operational planning is underlying to the approaches of Krajewska and Kopfer [21], Schwind et al. [29], and Berger and Bierwirth [2]. Groupage systems have first been introduced by Kopfer and Pankratz [16] with a focus on the cooperative framework of the system. Later approaches discuss procedures for the exchange of shipments. These exchanges were either conducted for individual shipments based on Vickrey auctions, or for bundled shipments based on combinatorial auctions. An approach referring to individual item Vickrey auctions is found in Berger and Bierwirth [2]. Since a common assumption is that profits can further be increased by bundling orders, most approaches consider the effects of exchanges based on combinatorial auctions. The approaches of Schönberger [27], Krajewska and Kopfer [21], and Schwind et al. [29] are examples for the pickup and delivery problem with less than truckload freights. All approaches consider at least three fulfillment options for requests: self-fulfillment, subcontracting and forwarding to one of the partners. Additionally, monetary transfer schemes for sharing the costs of collaborative fulfillment and sharing any collaborative profit are introduced in Krajewska and Kopfer [21] and Krajewska et al. [22]. A software system that re-allocates transportation requests among the profit centers of a company is introduced in Schwind et al. [29]. This approach uses combinatorial auctions for the assignment of clustered requests to profit centers. The authors assume perfect information for this setting and redistribute the commonly achieved revenues according to an exogenously determined activity index. The problem of including the reallocation of requests in a coalition is discussed in Berger and Bierwirth [2]. The reallocation is performed by considering all requests for fulfillment in one period and solving a multi-depot vehicle routing problem. The profit sharing assumes the existence of a perfectly informed central instance dividing the revenue among the participants.
66
H.W. Kopfer et al.
6.2.2 Characteristics and Advantages of Groupage Systems Being a coalition of cooperating forwarders, Groupage Systems face the same problems as the forwarders themselves. Additionally, Groupage Systems have to preserve the autonomy of the partners. Each of the participating forwarder interacts in a market where they receive demands for transportation from their customers which may be shippers or other forwarders. Groupage Systems focus on the operational planning of transportation request fulfillment – including the vehicle routing and scheduling for the own fleet of each partner, the forwarding of requests to subcontractors, and the opportunities of collaborative planning among the coalition partners. The discussion here is concentrated on the limit and the degree of autonomy in Groupage Systems. In contrast to the ongoing trend of concentration and unification of forwarding companies, most small and mid-sized enterprises want to stay autonomous and are looking for a close collaboration without building a centralized organizational unit. Thus, they want to cooperate within a Groupage System which provides occasions for the leveling of transportation capacities across voluntarily cooperating companies who remain legally and economically independent. Operational transport collaboration within Groupage Systems can create competitive advantages for the participating companies. The idea discussed for Groupage Systems is mainly based on a mutual adjustment of the operational planning of the partners of the coalition. This means, transportation requests are exchanged among the cooperating partners in order to improve profitability and service performance. In order to be successful a common exchange mechanism has to be found. Collaborative planning refers to the planning done jointly for all involved partners in the cooperation using the agreed exchange mechanism. Common to all considered approaches is the assumption that independent road haulage contractors enter only a certain share of their transportation requests to the collaborative planning. The remaining share of their requests is planned autonomously by each partner and this planning might even go further and include vertical cooperation with subcontractors. This leads to at least three modes of different related cost structures since subcontracting might be subject to more than one type of cost structure. The planning situation for the entire system is unlike the global approach for collaborative planning where plans are completely decided upon centrally. Rather, decentralized planning has to be installed where partners conduct their planning autonomously and only exchange limited information on a certain share of their customer requests. These shared requests are to be planned following the global objectives of the whole coalition. This kind of decentralized planning approach retains a higher degree of autonomy for the cooperating partners than the scenarios for horizontal cooperation typically found in practice. The underlying assumption of the suggested approach for distributed planning is that small and mid-sized enterprises have the desire to remain economically and organizationally autonomous and may be competitors when entering the cooperation. For collaborative planning this situation implies that not all information will be voluntarily provided and that it need not be excluded that partners might act strategically and provide incorrect information despite their desire to cooperate.
6
Limit and Degree of Autonomy in Groupage Systems
67
In order to find and select suitable exchange mechanisms for Groupage Systems it is necessary to define meaningful and significant criteria for the evaluation of promising mechanisms [3]. General criteria can be derived from microeconomic theory. These criteria are intended to be used for the evaluation of the efficiency of a system distributing bundled goods in an economy. The situation of planning in Groupage Systems is an extension of the theoretical case of perfect information towards asymmetric information and decentralized planning. Therefore, the criteria for the planning situation of autonomous partners in transport collaboration must be extended by the requirements of incentive and decentralized planning compatibility.
6.3 Prerequisites and Obstacles of Collaboration One of the main prerequisites for the success and sustainability of Groupage Systems is that each partner of the system must be satisfied with the collaboration on the short run and of course also on the long run. In particular this means, by being a member of the coalition, each partner should be better off than in the case of not being a part of the coalition. Additionally, each participant should have the feeling that they are not inferior compared to their partners. There are several prerequisites and obstacles for collaboration which can be divided in general impediments and special conflicts between single human decision makers [14]. Collaborative approaches are based on willingness, not on enforcement. The general impediments of collaboration mainly stem from the fact that sharing costs and revenues remains a highly controversial issue that can impede the collaboration process. Joining a collaboration means a high risk, because the involved companies share their individual own resources and partly subordinate to the coalition. Thus, the incentives to join the coalition must be relatively high. All partners have to be aware of attractive additional gains that will be achieved, and each of them has to accept potential risks that are arising by participating in the coalition. In order to establish and to keep alive the aspired cooperation among the partners, it is mandatory to demonstrate the potential benefits of cooperation to each partner, i.e. to show the expected additional profit. In order to prevent that taking part in the cooperation could eventually contribute to worsening the financial situation for some partners, mechanisms for compensational payments must be defined so that all participants achieve as good results as would be reached without cooperation. This requirement has to be preserved in a proper profit sharing concept which prescribes the kind of usage of the additional collaboration profit. The kind of usage of this profit constitutes the central point for the origin of conflicts. On the short-time level it is necessary that each single action of cooperation results in a positive effect for the acting partner. But also on the long-time level the collaboration profit must exceed the costs and the expectations of each partner, even if some partners act strategically. In some cases it is possible that conflicts with respect to profit utilization arise, although each partner is positively awarded for his or her participation within the coalition. This is usually based on subjective cognition of the partners. Such
68
H.W. Kopfer et al.
conflicts can for instance be triggered by the feeling of some partners that they find themselves to be inferior. In many of these situations the jealousy of partners plays an important role. An additional potential for conflicts is given by the question how services provided for the coalition should be recorded and measured. It is important for the partners that they can estimate the effects of their cooperative behavior in advance since incentives and decisions for collaborative activities should be coupled directly. Cooperation in a Groupage System means that persons in charge of several collaborating enterprises are supposed to interact in a predefined way. Such interactions can cause high conflict potentials on the managers’ and employees’ level. Additionally, establishing collaboration causes structural changes in the involved companies, because the core processes and organizational structures have to be unified in favor of standardized information flows. Regarding the conflict potentials, two types of conflicts can be identified. Conflicts can arise on the same hierarchy level between salesman and schedulers or on different hierarchy levels between salesman and schedulers on the one hand and their supervisors on the other hand. In case of different hierarchical levels, reluctance and fear for changes can intensify the relational conflicts between employees and managers. The antagonism of employees also causes so called “dual affiliation” conflicts for managers being associated with both, the other managers of partner companies (over the self-established collaboration structure) and with the subordinated persons (within the own company structure). The conflict of aims and the distributional conflict [20] typically arise among the teams of schedulers and salesmen of several partner companies. The schedulers have to compete for scarce profitable requests while having different proposals for their fulfillment. This completion may lead to jealousy effects on the employees’ level. Jealousy effects can be enforced by unequal proportions of power among companies. Schedulers and salesmen wish to dispose of non-profitable requests which are often also unattractive for the other partners. Collaboration processes are at risk to collapse because of an unfair information dispersal or insufficient transparency (e.g. if some information is held back by the employees of one company in the fight for attractive or against unattractive requests). Another challenge which amplifies the distributional conflict refers to trust. As the companies’ resources are used commonly (which causes additional common costs) and as the costs can be only generally assessed in advance, the gains are often not precisely predictable. Thus, the employees of the collaborating parties have to feel certain about the credibility and reliability of the partners. That is why, in order to achieve the minimum required level of adjustment among the partners and to allow the affordable exchange of information, mutual trust is inevitable.
6.4 Limits of Cooperation in Groupage Systems A Groupage System is built by a group of legally and economically independent enterprises who want to gain synergy effects by means of harmonized strategies and behaviors. The closeness of the relationship among the partners of the system
6
Limit and Degree of Autonomy in Groupage Systems
69
depends on the strength of cooperation which can vary between two extreme modes: (a) separate, non-cooperating enterprises which act in usual non-harmonized business connections on the one hand and (b) a fusion of enterprises on the other hand. The limits of cooperation in Groupage Systems are also given by the above mentioned extreme forms of business connections. The strength of cooperation can vary within these limits and is determined by the type and the intensity of resource exchange among the participating partners. The request exchange among the members of a Groupage System takes place during three successive phases. In the first phase, the participating partners can offer their desired options for request exchange and can announce these offers to all other partners within the coalition. This first phase is only used for making the options for request exchange visible. It refers both to the offering of requests that are put forward for exchange and to the bidding on requests. The bids indicate that a partner is willing to take a request or a bundle of requests on certain conditions. The second phase provides for a balancing between the interests of the coalition partners, which is necessary due to the expected incompatibility of the existing transfer options generated by the partners. In this sense the balancing has the effect of resolving the conflicting transfer wishes of the partners. The result of this balancing is an assignment of exchangeable requests to partners who have given bids on these requests. For the balancing process, respectively for conflict resolution, the interests of the entire coalition are relevant and these overall interests are estimated superior to the individual interests of the single coalition partners. This is because an assignment shall be achieved to make the total coalition profit as large as possible. This total coalition profit can then be distributed to the various partners involved. The individual interests of partners and possible disadvantages for single partners, resulting from the central assignment of requests can be compensated for by the distribution of the coalition profit. This distribution of the total coalition profit to the individual coalition partners will take place in the third phase, in the so-called profit-sharing phase. The profit-sharing phase is to guarantee that the collaboration renders positive benefits to each individual partner. The profit-sharing phase has two consistent goals. The first objective is to guarantee that the financial situation of each partner is not deteriorated as a result of the collaboration. The surplus of the coalition profit which is not needed for the fulfillment of the first objective is then divided according to the second objective. This has to be performed in such a manner, so that there is a reasonable incentive for individual partners to behave in a way which serves to benefit the entire coalition. The strength of collaboration influences the degree of autonomy of the partners in a Groupage System. Firstly, the strength of autonomy depends on the type of objects that are exchanged among the partners. The exchange of transportation resources – which means that they are left to each other execution planning – constitutes a closer connection and thus more collaboration strength than the exchange of transportation requests. Of course, the exchange of both will result to the closest form of collaboration. Secondly, the strength depends on the proportion of the amount of exchanged objects to the total amount of objects available in the coalition. The level of strength
70
H.W. Kopfer et al.
can vary within the lower and upper theoretical limits for the collaborative planning. In addition, there are different levels in between which identify specific levels of autonomy and will be discussed in Sect. 6.5. The lower theoretical limit for the strength of collaboration arises in the situation when all players act completely independently of each other. In this case there is no exchange of resources among the players and there is even no request exchange apart from usual business relations. The upper theoretical limit is obtained if all partners merge and submit to a central planning. In this way, the best theoretical planning results can be achieved, which is the global optimum for the overall planning problem. The centralization of the vehicle scheduling, however, would lead to a complete loss of autonomy of the coalition partners, which from an organizational point of view would be unacceptable for the partners. In addition, it is impossible to determine this global optimum out of technical reasons. First, the complexity and the computing power affordable for the central solution of the problem will be immense for larger Groupage Systems. Second, the fulfillment of the central planning would be very unstable and inflexible due to a huge number of requests that need to be scheduled at short notice, and because of the complicated relationships among the coalition partners. From a practical point of view there are lower and upper bounds for the degree of collaboration. These bounds are caused by practical issues and they are more restrictive than the above theoretical limits. A lower bound for the degree of collaboration is given by the requirement that the coalition should pay-off, which means that the benefits of cooperation should exceed the costs caused by cooperation. This implies that there should be a minimum amount of request exchange which compensates the fixed expenses of the coalition. On the other hand, the degree of collaboration is limited to an upper bound by the extent to which the partners are willing to share information, requests, and resources. The values of the lower and upper bounds for collaboration are strongly dependent on the compatibility of the partners and on the trust among the partners. Before a potential coalition is installed, it should be carefully checked whether the gap between the lower and upper bounds of the coalition leaves enough potential for a successful joint operation.
6.5 Levels of Collaboration in a Groupage System Following the goal of maximizing the value of the total profit of the coalition, it is generally recommended to seek a high intensity of collaboration within a Groupage System, because this leads to a significant expansion of the solution space for optimizing the request fulfillment. The intensity of interactions among partners of the coalition is determined by the extent of which they share or exchange objects, (i.e. by the type and amount of resources shared with each other, e.g. vehicles or drivers), by the amount of transportation requests offered for exchange, and the amount of requests actually exchanged among partners [17]. That is why the main criteria for the intensity of collaboration are given by the type of shared objects and the
6
Limit and Degree of Autonomy in Groupage Systems
71
proportion of shared requests. It is important to notice that we differ between effectively exchanged requests and potentially exchanged (i.e. offered) requests. The proportion of potentially exchanged requests is obtained in the first phase directly on amount of requests offered for exchange. In contrast, the proportion of actually transferred requests is determined at the end of the second phase. It depends on the amount of offered requests as well as on the amount and quality of the bids given for these offered requests. Additionally, it depends on the result of the optimization process determining the winners of the bidding process. Both, the amount of requests offered for exchange and the amount of bids depend to a large extent on the nature and intensity of the incentives granted by the profit-sharing model, according to which the partners will be rewarded for their behavior. Different levels of cooperation among freight forwarders are presented in [17]. The first level refers to the most intensive form of cooperation. In this form, vehicle scheduling is performed by global decision making. This means that all requests available in the total coalition are scheduled within a global plan by a central planning authority using all transport capacity available in the coalition. With respect to the operational planning and execution of transportation requests, this is equal to a fusion of the partners. The individual partners abandon their operational decision making competence with respect to their own requests and their own vehicles and leave this competence to the merger that they have commonly built. The central planning authority represented by this merger takes over the competence on all operational decisions including the decisions on the execution of all requests, and the decisions concerning the use of all vehicles. But this kind of cooperation differs from the fusion mentioned as theoretically upper level of cooperation in Sect. 6.4, since the partners still make their own decision on the tactical and strategic level, as for instance the decisions on the size and configuration of the own fleet. Even with respect to the operational field of vehicle routing they may strive for individual goals and follow specific restrictions that they want to be fulfilled. Of course, the representation of several individual objective functions and the inclusion of partnerspecific restrictions in a global optimization model is a challenging task from the mathematical point of view. On the second level, all transportation requests of all partners are gathered in an exchange pool and are offered for transfer. The transfer is made by an auction or any other pre-agreed mechanism for request exchange. This form of collaboration will result in the fact that Groupage partners will totally waive their right to decide on the planning and execution of their own transportation requests. Which partner will obtain which transportation request from the pool will be decided by the central decision procedure of the agreed exchange mechanism. All decisions referring to the further planning and execution of the obtained requests are then up to the acquiring partner of the coalition. It is typical for the second level of cooperation that, in contrast to the decisions on the assignment and execution of requests, the decisions on the use of the own resources remain at the owner of the resources. For this form of cooperation, the theoretically achievable solution quality for the vehicle routing and scheduling of all available requests usually is much worse than for the first form of collaboration, since the assignment of requests to performing forwarders is done
72
H.W. Kopfer et al.
in a decentralized way and strongly depends on the bids given by the partners. This means that unthrifty bids of some partners may have a negative influence on the construction of tours. Theoretically, the same solution quality can be achieved only if the exchange mechanism happens to lead exactly to the same allocation of requests to performing partners, which would also result in the simultaneous solution of the central approach pursued in the form of the first level of collaboration. The third level of collaboration is similar to the form on the second level but differs from it by the fact that each partner offers only a part of their requests for exchange. Each partner will decide independently which requests will remain private and which requests will be visible by putting them in the exchange pool. That means, each partner lays claim to decide on own requests and requests cannot be transferred, unless they explicitly are offered for exchange. It is crucial for this form of cooperation that the total coalition must succeed to achieve at least such an amount of collaboration profit that the profit exceeds the expenses for the cooperation including the increased transaction costs for collaboration. Only if there is a surplus of the collaboration profit on the transaction costs, it will be possible that the coalition partners will benefit from the cooperation and that incentives can be granted to the participating partners. The fourth level of cooperation is basically identical to the third level of collaboration. There is only a gradual distinction regarding the success or effectiveness of the coalition. On the fourth level, the benefit reached within the coalition due to the exchange of transportation requests is less than the costs, which are caused by the realization and maintenance of the cooperation. This may have different reasons, but in particular this may occur because the amount of transferred requests is so small that the resulting costs are not covered by the collaboration benefit. Of course, nobody would install such a coalition and it will certainly not survive for a longer time. But since it is not clear in advance whether a coalition will manage to generate a surplus of benefits over the resulting costs, such coalition exist, at least for some time. That is why the situation of loss-making coalitions should be analyzed, especially with the goal to decide whether the coalition could be saved or should be stopped. This decision will depend on the question whether the coalition suffers from structural deficiencies or from procedural deficiencies. In the latter case, measurements for the improvement of the processes involved in the request exchange can help to save the coalition.
6.6 Conclusions and Outlook The forms of joint planning and acting which have been discussed in Sect. 6.5 constitute different modes of cooperation on different levels with varying degrees of intensity. The autonomy of the involved partners is increasing from the first level to the last level and the potential savings reachable by cooperation decrease from level to level. In future research each level of cooperation and the associated forms of joint planning should be investigated in detail. In this research the following
6
Limit and Degree of Autonomy in Groupage Systems
73
characteristics of the considered collaboration forms should be analyzed: (a) the degree and the consequences of economic and legal dependence, (b) proximity to customers, (c) development of suitable models and algorithms for centralized and decentralized vehicle routing and scheduling, (d) nature and extent of the transmitted sensitive information among the partners, (e) privacy and transparency within the entire coalition (e.g. with respect to the set of available requests, the available fleet capacity, the running costs of the fleet of the individual partners, and the freight rate paid by the customers of individual partners), (f) customer protection, (g) achievable solution quality of the generated transportation plans, (h) transaction costs, and (i) requirements for an appropriate profit-sharing model. Different levels of cooperation require different and adequate exchange mechanisms. The major challenge for future research on exchange mechanisms lies in the computational complexity of the underlying problems to be solved for evaluating transportation requests and reallocating them by assigning them to the collaborating partners. Further, the situation of sellers has not been considered so far but needs to be addressed [18]. Every participant may take the position of seller and buyer at the same time. It is a rational goal for the seller to gain additional profit by selling transportation requests. However, if this additional profit becomes too large, it may contradict an efficient solution. Additionally, sellers might not necessarily provide correct information on transportation requests and thus endanger the calculation of efficient transfer prices. In our contribution we have analyzed the chances and the changes of operational transportation planning that result from adding the option of horizontal cooperation among the members of a closed group of forwarders or carriers. A deeper analysis of the various forms of collaboration and their characteristics enables the assessment and the systematic configuration of coalitions on the basis of their specific advantages and disadvantages.
References 1. Baganha M, Cohen M (1998) The stabilising effect of inventory in supply chains. Oper Res 46:572–583 2. Berger S, Bierwirth C (2010) Solutions to the request reassignment problem in collaborative carrier networks. Transport Res E 46:627–638 3. Bloos M, Schönberger J, Kopfer H (2009) Supporting cooperative demand fulfillment in supply networks using autonomous control and multi-agent-systems. In: Fischer S, Maehle E, Reischuk R (eds) INFORMATIK 2009. Im focus das leben. GI-edition lecture notes in informatics, 3590–3604 4. Carter C, Jennings M (2002) Social responsibility and supply chain relationships. Transport Res E 38:37–52 5. Cruijssen F, Cools M, Dullaert W (2005) Horizontal cooperation in logistics: opportunities and impediments. Transport Res E 43:129–142 6. Cruijssen F, Salomon M (2004) Empirical study: order sharing between transportation companies may result in cost reductions between 5% to 15%. CentER Discussion Paper 2004–80, Faculty of Economics and Business Administration, Tilburg University, The Netherlands
74
H.W. Kopfer et al.
7. Ergun Ö, Kuyzu G, Savelsbergh MWP (2007) Shipper collaboration. Comput Oper Res 34:1551–1560 8. Fernandez F, Hinojosa M, Puerto J (2004) Multi-criteria minimum cost spanning tree games. Eur J Oper Res 158:339–408 9. Fu Y, Piplani R (2004) Supply-side collaboration and its value in supply chains. Eur J Oper Res 152:281–288 10. Gentry J (1993) Strategic alliances in purchasing: transportation is the vital link. Int J Purch Mater Manag 29:11–17 11. Huang Z, Li S (2001) Co-op advertising models in manufacturer-retailer supply chains: a game theory approach. Eur J Oper Res 135:527–544 12. Jørgensen S, Taboubi S, Zaccour G (2003) Retail promotions with negative brand image effects: is cooperation possible? Eur J Oper Res 150:395–405 13. Kilger C, Reuter B (2002) Collaborative planning. In: Stadtler H, Kilger C (eds) (2002). Supply chain management and advanced planning, pp 223–241. Springer, Berlin-HeidelbergNew York 14. Kopfer HW, Kopfer H (2009) Hemmnisse horizontaler Kollaboration in der Speditionsbranche. In: Ivanov D, Meinberg U (eds) Logistics and supply chain management: modern trends in Germany and Russia, pp 354–362. Cuvillier Verlag, Göttingen 2009 15. Kopfer H, Krajewska M-A (2007) Praktische aspekte der kollaborativen auftragsdisposition. In: Koschke R, Herzog O, Rödiger K-H, Ronthaler M (eds) INFORMATIK 2007. Informatik trifft Logistik. GI-edition lecture notes in informatics, pp 66–69 16. Kopfer H, Pankratz G (1999) Das groupage-problem kooperierender verkehrsträger. In: Kall P, Lüthi H-J (eds) Proceeding of OR 98, pp 453–462. Springer, Berlin, Heidelberg, New York 17. Kopfer H, Kopfer HW, Wang X (2010) Grad und grenzen kollaborativer tourenplanung. In: Ivanov D, Lukinskiy V, Sokolov B, Kaeschel J (eds) Logistik and supply chain management, pp. 140–147. Deutsch-Russische Perspektiven, St. Petersburg 18. Kopfer H, Schönberger J, Bloos M (2010) Groupage systems-collaborative request fulfilment in road haulage: a procedural view. In: Schönberger R, Elbert R (eds) Dimensionen der Logistik: Funktionen, Institutionen und Handlungsebenen, pp 787–804. Gabler, Wiesbaden 19. Kopfer H, Wang X (2009) Combining vehicle routing with forwarding – extension of the vehicle routing problem by different types of sub-contraction. J Korean Inst Ind Eng (JIKIIE) 35(1):1–14 20. Krajewska M-A (2008) Potentials for efficiency increase in modern freight forwarding. Gabler Edition Wissenschaft, Wiesbaden 21. Krajewska M-A, Kopfer H. (2006) Collaborating freight forwarding enterprises – request allocation and profit sharing. OR Spectrum 28(3):301–317 22. Krajewska M-A, Kopfer H, Laporte G, Ropke S, Zaccour G (2008) Horizontal cooperation of freight carriers: request allocation and profit sharing. J Oper Res Soc 59:1483–1491 23. Krus L, Bronisz P (2000) Cooperative game solution concepts to a cost allocation problem. Eur J Oper Res 122:258–271 24. Li S, Huang Z, Zhu J, Chau P (2002) Cooperative advertising, game theory and manufacturerretailer supply chains. Omega 30:347–357 25. Matsubayashi N, Umezawa M, Masuda Y, Nishino H (2005) A cost allocation problem arising in hub-spoke network systems. Eur J Oper Res 160:821–838 26. Pfohl H-C (2004) Logistikmanagement. Konzepte und Funktionen. Springer, Berlin 27. Schönberger J (2005) Operational freight carrier planning. basic concepts, optimization models and advanced memetic algorithms. GOR Publications, Springer, Berlin, Heidelberg, New York 28. Schönsleben P (2000) With agility and adequate partnership strategies towards effective logistics networks. Comput Ind 42:33–42 29. Schwind M, Gujo O, Vykoukal J (2009) A combinatorial intra-enterprise exchange for logistics services. Inform Syst E-Bus Manag 7:447–471 30. Slats P, Bhola B, Evers J, Dijkhuizen G (1995) Logistic chain modelling. Eur J Oper Res 87:1–20
6
Limit and Degree of Autonomy in Groupage Systems
75
31. Slikker M, Fransoo J, Wouters M (2005) Cooperation between multiple news-vendors with transshipments. Eur J Oper Res 167:370–380 32. Tate K (1996) The elements of successful partnership. Int J Phys Distrib Logist Manag 26:7–13 33. Yue J, Austin J, Wang M, Huang Z (2006) Coordination of cooperative advertising in a twolevel supply chain when manufacturer offers discount. Eur J Oper Res 168:65–85
Chapter 7
The Interaction Effort in Autonomous Logistics Processes: Potential and Limitations for Cooperation Arne Schuldt, Jan Ole Berndt, and Otthein Herzog
7.1 Introduction The complexity, dynamics, and distribution of logistics processes are major challenges for supply network management. The paradigm of autonomous control in logistics [32] addresses these challenges by delegating process control to the participating logistics entities. Autonomous logistics can be implemented with intelligent software agents that represent the logistics objects and act on their behalf [26,30,33]. Decentralising decision-making enables a problem decomposition which, in turn, reduces the computational effort. Every entity has to incorporate only its own parameters plus those of cooperating entities. Simultaneously, occurring dynamics can be dealt with locally. While problem decomposition reduces the computational effort, another challenge arises: Individual entities can hardly satisfy their objectives on their own. Often, cooperation is necessary. Appropriate organisational structures being established on demand facilitate the required coordination. Simultaneously, organisation can decrease the interaction effort for process control. However, forming these structures and performing activities cooperatively also induces a certain interaction effort. Designers of autonomous logistics processes must trade off both aspects against each other to prevent that the decrease in computational effort is outweighed by the increased interaction effort. This chapter contributes a thorough investigation of the interaction effort for different types of cooperation in autonomous logistics. The remainder of this chapter is structured as follows: Sect. 7.2 introduces a process that will serve as a recurring example. Based on that foundation, Sect. 7.3 examines the potential for cooperation in autonomous logistics processes. To this end, different types of cooperation are distinguished and analysed. Subsequently, Sects. 7.4 and 7.5 describe how the respective cooperation mechanisms can be implemented. Section 7.6 examines the interaction effort for coordinating autonomous logistics A. Schuldt (B), J.O. Berndt, and O. Herzog TZI IS, Universität Bremen Am Fallturm 1, 28359 Bremen, Germany e-mail:
[email protected],
[email protected],
[email protected] M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_7,
77
78
A. Schuldt et al.
entities. A particular focus lies on limitations for cooperation and how they can be dealt with. This helps narrow an appropriate degree [31] for autonomous control in logistics.
7.2 Investigated Process Onward carriage of sea containers serves as a recurring example throughout the remainder of this chapter. The respective real-world process and its transition to autonomous logistics have been thoroughly investigated [26]. Onward carriage pertains to intermodal transport of sea containers which can be divided into three parts. During pre carriage, the container is transported from the sender to the port of loading. Subsequently, the seaborne transport by container vessel is conducted. The last step is onward carriage from the port of discharge to the receiver of the cargo. The examined real-world process covers two logistics functions, namely transport and storage. The requirements for the logistics objects are as follows: 1. Autonomous shipping containers have to choose a storage facility at which their cargo can be received and stored. 2. Autonomous shipping containers have to choose a transport service provider for transport to the selected storage facility. For both tasks, it is necessary to find service providers that are capable of handling the respective cargo. Furthermore, these service providers must have sufficient capacity. Finally, the cheapest matching and available service provider should be chosen. Apart from these criteria for individual objects, additional group-related requirements have to be considered: 1. Similar goods should be stored together in the same storage facility whenever possible. 2. Multiple containers should be transported together on means of mass transport whenever possible.
7.3 Potential for Cooperation in Autonomous Logistics On the one hand, satisfying Requirements 1 and 2 identified in the preceding section exceeds the capabilities of individual logistics entities. On the other hand, the high degree of spatial distribution of the logistics objects prevents a central entity from taking over coordination tasks [26]. Consequently, autonomous logistics entities must coordinate themselves by establishing adequate organisational structures. Cooperative problem solving [34] can be divided into four steps (Fig. 7.1). Initially, the logistics entities must recognise a potential for cooperation because cooperation is not worthwhile otherwise (Recognition). Having recognised this potential, they
7
The Interaction Effort in Autonomous Logistics Processes
Recognition
Team Formation
Plan Formation
79
Team Action
Fig. 7.1 The four steps of the cooperative problem solving process. As a precondition, logistics entities must recognise a potential for cooperation. Subsequently, they form both a team and a joint plan before they actually act jointly
can form teams that are beneficial with regard to their respective goals (Team Formation). Afterwards, the team members agree upon a joint plan (Plan Formation). Finally, the actual execution of the joint plan takes place as the last process step (Team Action). Each of these steps can fail, usually leading to the entities having to return to the preceding step. The potential for cooperation as a prerequisite for joint action can be identified as follows [34]. Firstly, a logistics entity must recognise that there exists some team of entities that can jointly achieve the intended goal. Secondly, the entity must have a reason for acting in a team. This holds if it cannot achieve its goal in isolation or if there is a goal conflict for all actions it could perform to achieve the goal. In autonomous logistics, the potential for cooperation particularly manifests in the following aspects: 1. Complementing insufficient individual capabilities. 2. Increasing the resource utilisation efficiency. Sections 7.3.1 and 7.3.2 examine both aspects in more detail with a particular focus on applications in autonomous logistics.
7.3.1 Complementing Insufficient Individual Capabilities In autonomous logistics, general cargo units are expected to plan and schedule their path through the logistics network themselves. However, these autonomous logistics entities can usually not accomplish their respective tasks on their own. Instead, they have to delegate operations to suitable service providers. As an example from the onward carriage process (Sect. 7.2), an autonomous shipping container cannot move from its source to its sink on its own but has to be transported. Thus, in order to satisfy its objective, the container must be enabled to form a team with one or more means of transport. Team formation helps autonomous logistics entities complement their insufficient individual capabilities by utilising resources provided by other entities. However, choosing the right partners for cooperation is a difficult task. This choice depends not only on the service type and on spatiotemporal constraints [28], but also on quality criteria like service provider responsiveness and reliability. The former criteria express situational conditions for making a decision. The latter ones describe expectations towards potential cooperation partners and reflect business relationships as well as experience from previous interaction activities.
80
A. Schuldt et al.
The dynamics of logistics processes prevent design time evaluation and optimisation of potential cooperations in terms of flexibility and robustness. The following reasons require online formation and adaption of cooperative activities in autonomous logistics: 1. Scheduled operations can fail 2. Demands and capabilities can vary 3. Service providers and consumers can enter and leave the system The first two reasons lead to the need for replanning and reallocating resources. The third reason is due to the openness of such systems [3]. Service providers and consumers entering and leaving the logistics system aggravate the inherent dynamics of logistics processes. They induce further changes in availability, demand, prices, and requirements of products and services. To summarise, there is a potential for complementing insufficient individual capabilities in autonomous logistics. However, this kind of cooperation cannot be defined at design time but must evolve during runtime. The dynamics of logistics processes evoke the need for each autonomous entity to constantly reconsider and modify its cooperative relationships to other entities in order to effectively reach its objectives.
7.3.2 Increasing the Resource Utilisation Efficiency Individual autonomous logistics entities can frequently not meet minimum utilisations required by service providers. To reiterate the onward carriage scenario (Sect. 7.2), transport by train is preferable over transport by truck due to lower transport rates. However, the train has a certain minimum utilisation consumers must accept when requesting its service. Transporting an individual container by train alone is thus more expensive than transporting it by truck. Consequently, in autonomous logistics multiple shipping containers can benefit from coordinating themselves for joint resource utilisation. Figure 7.2 depicts a graph that derives this potential for cooperation analytically. One coordinate axis refers to the number of service providers employed which ranges from 1 to 1;000. The other axis refers to the lot size of each provider, thereby also ranging from 1 to 1;000. The graph depicts the actual overall utilisation induced by 1;000 service consumers, whereby the consumers are assumed to be equally distributed to the service providers. The graph clearly shows that the need for restricting the number of service providers increases with the lot sizes. If the lot size is 1, it does not make any difference how many service providers are employed. However, with increasing lot sizes also the overall utilisation increases. In case of the lot size being 1;000, the 1;000 service consumers employing 1;000 different providers induce a total utilisation of 1;000;000 although 1;000 would suffice. This means that the higher the lot size of
7
The Interaction Effort in Autonomous Logistics Processes
81
Utilisation
Utilisation
1200000
1200000
1000000
1000000
800000
800000
600000
600000
400000
400000
200000
200000
0 1
0
100
200
300
400
Providers
500
600
700
800
900 1000 1
100
200
300
400
500
600
700
800
1000 900
Lot Size
Fig. 7.2 The potential for cooperation of 1; 000 service consumers by joint utilisation of logistics resources. The higher the lot size of the logistics service providers, the higher is the benefit from jointly utilising only a few selected providers
the logistics service providers, the higher is the benefit from jointly utilising only a few selected providers. Hence, there is a massive potential for cooperation.
7.4 Using Resources Effectively As elaborated in the previous section, finding appropriate cooperation partners that complement own insufficiencies is crucial for effective process control. Section 7.4.1 derives relevant requirements for this task and examines related work. Subsequently, Sect. 7.4.2 describes how effective resource utilisation can be implemented. The particular focus is on how autonomous logistics service consumers can find suitable partners for cooperation among the service providers.
7.4.1 Requirements and Related Work The logistics domain is shaped by a high degree of both complexity and dynamics and the multiplicity of criteria for the choice of potential cooperation partners
82
A. Schuldt et al.
(Sect. 7.3.1). These properties lead to the following requirements for an interaction of autonomous logistics entities facilitating effective resource utilisation: 1. 2. 3. 4.
Decentralisation Autonomy Adaptivity Reliability
The demand for choosing suitable cooperation partners in a decentralised way (Requirement 1) is a direct consequence of the decentralised setting in autonomous logistics. Likewise, the autonomy of the logistics entities in their choice of cooperation partners must be preserved (Requirement 2). In order to enable effective cooperation of these entities, channels and modes for their interaction have to be defined. To this end, a wide variety of different paradigms for interaction structures has been proposed [13], ranging from strict hierarchies [17] to negotiation-based mechanisms [4, 29]. These organisational structures defining the interaction competencies of autonomous entities are chosen at the design stage of a system and thus determine its runtime behaviour. However, cooperation processes must be able to react to changes caused by the dynamics of logistics, i.e., cooperative relationships have to be reconsidered continuously (Requirement 3). Adjusting the patterns of interaction along the spectrum of organisational approaches between delegation and negotiation helps cope with these conditions and attain scalability of cooperation [23, 24]. However, it remains unclear under which conditions which organisational structure is suitable for ensuring effective cooperation. Not only the dynamics of logistics, but also the capabilities and the behaviour of other entities affect the effectiveness of cooperative activities [19]. This impacts the reliability of the cooperation outcome (Requirement 4). In an open system, the outcome cannot be determined beforehand, but must be assessed from observations during runtime of the system. Based on experience from such observations, it is necessary to model expectations concerning the behaviour [3, 18] of other logistics entities that allow for estimating the outcome of cooperative activities.
7.4.2 Emergent Interaction Patterns Inspired by Luhmann’s sociological theory of communication systems [15], feeding the aforementioned expectations back into the decision-making process results in a control loop enabling the emergence of effective cooperation in autonomous logistics [2]. This process consists of three consecutive steps that are executed repeatedly: 1. Estimate the outcome of potential cooperations from past observations. 2. Choose cooperative activities according to the expected results. 3. Memorise actually observed operations outcome. In this approach, expectations of the results of potential cooperations with other logistics entities are based on the entities storing their observations in a memory.
7
The Interaction Effort in Autonomous Logistics Processes
83
The outcome of joint activities is therefore estimated using the memory entries describing past observations of similar operations. Choosing its cooperations in accordance with their expected results, an entity will then be able to observe their actual outcome and to store these observations in its memory again. Therewith, the control loop is closed, as the updated memory entries are used to calculate the expected outcome of further activities. Thus, logistics entity relationships emerge from interaction processes while guiding the choice of subsequent operations. Regarding the properties of the logistics domain, the advantage of these selforganising activity patterns is that the range of potential cooperations can be narrowed to those being expected to have the most promising results. Hence, logistics entities become able to learn best practices in cooperatively utilising resources and capabilities which leads to a gradually increasing performance of the whole system [2]. Moreover, being based on experience, the choice of a logistics service provider as a partner for joint activities not only reflects its reliability according to past observations, but also allows for behavioural adaptions in case of changing conditions: If the actual observations do no longer meet the expected results, the logistics entities will reconsider and adapt their cooperative relationships in order to re-enable effective utilisation of distributed resources.
7.5 Using Resources Efficiently Joint action has been identified important not only for effective, but also for efficient resource utilisation (Sect. 7.3.2). The required coordination of autonomous logistics entities with similar objectives or properties can be accomplished with respective team formation mechanisms. Section 7.5.1 derives requirements for such mechanisms and examines the applicability of previous approaches. Afterwards, Sect. 7.5.2 describes how team formation for joint resource utilisation can be solved in autonomous logistics.
7.5.1 Requirements and Related Work Previous work on team formation formalises organisations of autonomous entities [6, 7, 10] as well as the internal states of these entities during team formation [5,34]. By contrast, less effort has been spent on particular interaction protocols for team formation. For the design of such interaction mechanisms, general design principles should be considered [21, 22]. In addition, the intended team formation mechanism for autonomous logistics should meet the following requirements: 1. 2. 3. 4. 5.
Decentralisation Genericness No prior knowledge Unique teams Flexible teams
84
A. Schuldt et al.
To reiterate the onward carriage process (Sect. 7.2), autonomous control demands that shipping containers with similar properties and objectives act jointly. Team formation of autonomous shipping containers is an inherently distributed problem (Requirement 1). That is, no individual container has a centralised perspective that incorporates all parameters. Consequently, previous cluster algorithms like k-means [16] are not applicable. Approaches to distributed clustering can be found in wireless sensor networks [1]. Usually, these approaches focus on spatial data and make thus implicit assumptions on the environment [11]. This means that they are not applicable for the generic descriptors required in autonomous logistics (Requirement 2). Peer-to-peer approaches [20] do not exhibit this restriction. By contrast, each logistics entity has an arbitrarily chosen set of other entities it is initially connected with. Autonomous entities can then exchange their direct partners by more similar ones. This approach, however, contradicts the requirements that the required prior knowledge should be minimised (Requirement 3). In particular, there is no natural choice for initial peers in autonomous logistics. A promising candidate for implementing team formation is the contract net [29] interaction protocol. Although it is well-suited for complementing capabilities (Sect. 7.3.1), its applicability to efficient process control is limited. This can be explained as follows: In terms of onward carriage (Sect. 7.2), shipping containers with similar properties and objectives should form one team (Requirement 4). For instance, this means there should only be one team of packages for each pair of origin and destination. Of course, each team of consumers can intentionally employ multiple trucks if necessary. Unique teams ensure the highest utilisation of the logistics resource employed. In other words, the requirement for unique teams means that autonomous logistics entities group themselves by an equivalence relation with respect to one property or a set of properties. Nevertheless, different properties and thus varying teams may exist for different purposes, e.g., transport and storage service allocation. Closely related to unique teams is the requirement for flexible teams (Requirement 5), i.e., members must be able to join after a team has been initially established. With the contract net protocol, team managers would have to continuously advertise their team to potential new members. Despite of efficiency issues, the question remains open what might be an appropriate frequency for such announcements.
7.5.2 Team Formation Interaction Protocol Team formation for efficient resource utilisation includes the following roles: the initiating participant, already existing team managers, and a directory that administers the list of the current managers. New entities participating in team formation, initially register themselves as new team managers with the directory service. This behaviour is based on the optimistic assumption that no current team manager matches its properties for team formation. To check whether this assumption is true, all team managers are provided with the
7
The Interaction Effort in Autonomous Logistics Processes
85
properties of the new participant. The list of team managers can be requested from the directory. All communication with the team managers is conducted in parallel. It has three possible results [27]: 1. No team manager identifies a match. Then, the optimistic assumption turns out to be right. Consequently, the new participant actually becomes a new team manager. 2. One team manager determines a match. Then, the initial assumption was wrong. Consequently, the new participant deregisters from the directory and joins the matching team instead. 3. Multiple team managers determine a match. This may happen if multiple logistics entities have concurrently registered with same properties. This can be resolved by assigning an individual registration timestamp to each team manager. Then, all entities including the superfluous team managers can join the initial team. An alternative approach to optimistic behaviour is conservative interaction as performed in a previous version of the protocol [28]. In that case, new participants first contact all existing team managers before registering themselves with the directory. Still, it cannot be prevented that multiple logistics entities with same properties register as team managers because all entities act concurrently. To resolve this redundancy, all existing team managers must thus be contacted again after the registration process. The optimistic procedure thus has the advantage that team managers only have to be contacted once. Both approaches have in common that they work in a decentralised way. They make neither special assumptions regarding the team descriptors applied nor special requirements regarding prior knowledge about the system. Furthermore, they enable forming unique and flexible teams as required for efficient resource utilisation.
7.6 Appropriate Degree for Autonomous Control As an intermediary result, Sects. 7.4 and 7.5 have identified methods for allocating logistics resources with autonomous control both effectively and efficiently. Autonomous logistics is motivated by the decrease in computational effort achieved by problem decomposition. The higher the degree of decentralisation, however, the higher the need for coordination. That is, the interaction effort increases with the number of participating autonomous logistics entities. Therefore, it is important to identify limitations and thereupon an appropriate degree at which autonomous control should be conducted. Section 7.6.1 examines the interaction effort for complementing insufficient individual capabilities effectively. This investigation leads to the insight that joint action based on joint objectives or properties does not only increase the resource utilisation efficiency, but also reduces the interaction effort. Section 7.6.2 then addresses the interaction effort for team formation, an important prerequisite for adjusting the degree of autonomous control by joint action.
86
A. Schuldt et al.
7.6.1 Interaction Effort for Resource Utilisation For an autonomous logistics entity, choosing an appropriate partner for cooperation is constrained by the type of service required, by spatiotemporal criteria, and by free capacities as well as by the expected reliability of each service provider (Sect. 7.3.1). These aspects help narrow the set of potential cooperation partners from all logistics service providers to those that are actually able to fulfil the given demands. However, the properties of autonomous logistics entities change over time caused by the dynamics of logistics processes. For instance, means of transport change their spatial position and may be loaded or empty. As another example, storage facilities have fluctuating utilisation rates. Therefore, finding a suitable service provider requires an autonomous entity to consider all potentially matching ones. The effort needed for effectively utilising resources in autonomous logistics is generally linear to the number of providers of a respective resource for each single logistics entity intending to utilise it. Consequently, the overall effort is equal to the number of service providers n times the number of service consumers m. The asymptotical complexity for resource utilisation is thus O.nm/ D O.n2 /. As the effort needed for resource utilisation depends directly on the number of interacting entities, that very number is the key determinant for an appropriate degree of autonomous control. If the required amount of interaction processes exceeds the computational gain achieved by problem decomposition, it renders the chosen degree of autonomy unsuitable. Yet, joint action of logistics entities pursuing equivalent objectives allows not only for efficient resource usage (Sect. 7.5), but also helps reduce the number of actually interacting participants. In a team of autonomous logistics entities, only one of them represents the whole team. Thus, the number of interacting entities is reduced by the ratio of entities to teams (Fig. 7.3). Hence, cooperation of autonomous entities with similar objectives allows for retaining the computational advantages of problem decomposition as well as the scalability of autonomous control in logistics. In addition to the reduction of the interaction effort for effective resource usage, teams of logistics entities benefit from the collective experience of their individual members [14]. Exchanging memorised information about cooperation reliability (Sect. 7.4.2) allows for the fast emergence of effective interaction structures by deriving expectations from the team’s accumulated observations. Thus, cooperation also increases the ability of autonomous entities to react to changing conditions by using shared expectations for reducing the number of observations necessary for each single entity. As a preliminary conclusion, the appropriate degree for autonomous control depends on the reduction of computational effort on the one hand compared to the increase of interactions required for coordination on the other. The latter furthermore is determined by the total number of participating entities and the potential for joint pursuit of similar goals. Nevertheless, while cooperation helps confine the interaction effort and even improves the adaptivity of an autonomous logistics system, the formation of a team of entities itself requires further interaction. Therefore,
7
The Interaction Effort in Autonomous Logistics Processes
87
Interaction Effort (Ratio)
Interaction Effort (Ratio)
1.2
1.2
1.0
1.0
0.8
0.8
0.6
0.6
0.4
0.4 0.2
0.2 0 1000
900
0
800
700
600
Teams
500
400
300
200
100
1
1
100
200
300
400
500
600
700
800
900
1000
Agents
Fig. 7.3 The potential for decreasing the interaction effort for resource utilisation by cooperation. A high average number of entities per team leads to a low ratio of actually required interaction effort
the next section examines the complexity of team formation in order to facilitate determination of an appropriate degree of autonomous control in logistics.
7.6.2 Interaction Effort for Team Formation Determining the interaction effort for team formation by directory is not trivial. In principle, each of the n participants in team formation has to contact all m team managers (Sect. 7.5.2). This means that the interaction complexity is O.nm/ D O.n2 /. For most applications, however, it holds that m n. The actual interaction effort [27] depends not only on the pure amount of team managers and members. Instead, it also depends on the time at which participants become team managers. In the course of time, the number of team managers increases and therewith also the number of entities new participants have to coordinate with. Figure 7.4 depicts the total interaction effort for a scenario of 1; 000 entities which form between 1 and 1; 000 teams. The graph counts the total number of interactions until the appearance of the respective entity. The lowest interaction effort arises if only one team exists which is joined by all participants. The highest effort arises if every participant forms its own team.
88
A. Schuldt et al.
Interactions
Interactions
600000
600000
500000
500000
400000
400000
300000
300000
200000
200000
100000
100000
0 1000
900
0
800
700
600
500
Teams
400
300
200
100
1
1
100
200
300
400
500
600
700
800
900
1000
Agent
Fig. 7.4 The interaction effort for team formation of 1,000 entities based on a directory service. The interaction effort increases with the number of teams established
A limitation of the proposed team formation interaction protocol (Sect. 7.5.2) is reached when the interaction effort exceeds the capabilities of either the individual entities or the whole logistics system. Two alternatives exist for resolving this situation: 1. The number of participating entities can be restricted, e.g., based on spatial regions of relevance [9]. 2. Instead of a directory, a broker can be employed, thereby decreasing both the degree of decentralisation and the interaction effort [25]. To summarise, team formation helps reduce the interaction effort for resource utilisation. It particularly enables choosing an appropriate degree at which autonomous control is applied. Team formation itself requires a certain interaction effort. This can be approached by carefully choosing both the participants and the interaction mechanisms.
7.7 Conclusion Finding an appropriate degree at which autonomous control is applied is one of the key tasks in designing autonomous logistics processes. On the one hand, problem decomposition helps reduce the computational effort. On the other hand,
7
The Interaction Effort in Autonomous Logistics Processes
89
the interaction effort increases with decentralisation because the participants must coordinate each other. Cooperation of autonomous logistics entities is thus an important foundation for autonomous control. To narrow the optimal degree for autonomous control, the potential for cooperation in autonomous logistics is identified. Cooperation is a prerequisite for both effective resource allocation by matching complementing capabilities and efficient resource utilisation. Methods for implementing both occurrences of cooperation are presented. Apart from their potential, these methods are also examined regarding their limitations in order to derive an appropriate degree for autonomous control. It turns out that the interaction effort can usually be reduced by joint action. To this end, autonomous logistics entities should cooperate based on joint objectives or properties. However, team formation itself also induces interaction effort. Therefore, it is important to choose team formation mechanisms carefully based on the intended application.
References 1. Akyildiz IF, Su W, Sankarasubramaniam Y, Cayirci E (2002) Wireless sensor networks: a survey. Comput Network 38(4):393–422 2. Berndt JO (2011) Self-organizing supply networks: autonomous agent coordination based on expectations. In: ICAART 2011, INSTICC Press, Rome, Italy 3. Brauer W, Nickles M, Rovatsos M, Weiß G, Lorentzen KF (2002) Expectation-oriented analysis and design. In: AOSE 2001, Springer, Montreal, QC, Canada, pp 226–244 4. Collins J, Youngdahl B, Jamison S, Mobasher B, Gini M (1998) A market architecture for multi-agent contracting. In: Agents 1998, ACM Press, Saint Paul, MN, USA, pp 285–292 5. Dignum R, Dunin-Keplicz B, Verbrugge R (2000) Agent theory for team formation by dialogue. In: ATAL 2000, Springer, Boston, MA, USA, pp 150–166 6. Ferber J, Gutknecht O (1998) A meta-model for the analysis and design of organizations in multi-agent systems. In: ICMAS 1998, IEEE Computer Society, Paris, France, pp 128–135 7. Fischer K, Schillo M, Siekmann JH (2003) Holonic multiagent systems: a foundation for the organisation of multiagent systems. In: HoloMAS 2003, Springer, Prague, Czech Republic, pp 71–80 8. Fischer K, Florian M, Malsch T (eds) (2005) Socionics. Scalability of complex social systems. Springer, Heidelberg, Germany 9. Gehrke JD (2009) Evaluating situation awareness of autonomous systems. In: Madhavan R, Tunstel E, Messina E (eds) Performance evaluation and benchmarking of intelligent systems. Springer, Heidelberg, Germany, pp 93–111 10. Hannoun M, Boissier O, Sichman JS, Sayettat C (2000) MOISE: an organizational model for multi-agent systems. In: IBERAMIA-SBIA 2000, Springer, Atibaia, Brazil, pp 156–165 11. Heinzelman WR, Chandrakasan A, Balakrishnan H (2000) Energy-efficient communication protocol for wireless microsensor networks. In: HICSS 2000, IEEE Computer Society, Maui, Hawaii, vol 8, pp 8020–8029 12. Hülsmann M, Windt K (eds) (2007) Understanding autonomous cooperation and control in logistics. Springer, Heidelberg, Germany 13. Horling B, Lesser V (2005) A survey of multi-agent organizational paradigms. Knowl Eng Rev 19(4):281–316, URL http://mas.cs.umass.edu/paper/366 14. Langer H, Gehrke JD, Hammer J, Lorenz M, Timm IJ, Herzog O (2006) A framework for distributed knowledge management in autonomous logistic processes. Int J Knowl Base Intell Eng Syst 10(4):277–290
90
A. Schuldt et al.
15. Luhmann N (1995) Social systems. Stanford University Press, Stanford 16. MacQueen J (1967) Some methods for classification and analysis of multivariate observations. In: 5th Berkeley Symposium on Mathematical Statistics and Probability, University of California Press, Berkeley, CA, USA, pp 281–296 17. Montgomery TA, Durfee EH (1993) Search reduction in hierarchical distributed problem solving. Group Decis Negot 2:301–317 18. Nickles M, Weiss G (2005) Multiagent systems without agents – mirror-holons for the compilation and enactment of communication structures. In: Fischer K, Florian M, Malsch T (eds) (2005) Socionics. Scalability of complex social systems. Springer, Heidelberg, Germany, pp 263–288 19. Nickles M, Rovatsos M, Weiss G (2005) Expectation-oriented modeling. Eng Appl Artif Intell 18(8):891–918, DOI DOI:10.1016/j.engappai.2005.05.002, URL http://www.sciencedirect. com/science/article/B6V2M-4GJK86T-1/2/74b3c8947f7310792c5a5cb3235a091d 20. Ogston E, Vassiliadis S (2001) Matchmaking among minimal agents without a facilitator. In: Agents 2001, ACM, Montreal, QC, Canada, pp 608–615 21. Rosenschein JS, Zlotkin G (1994) Rules of encounter: designing conventions for automated negotiation among computers. MIT, Cambridge, MA, USA 22. Sandholm TW (1999) Distributed rational decision making. In: Weiss G (ed) Multiagent systems. A modern approach to distributed artificial intelligence, MIT, Cambridge, MA, USA, pp 201–258 23. Schillo M, Spresny D (2005) Organization: the central concept for qualitative and quantitative scalability. In: Fischer K, Florian M, Malsch T (eds) (2005) Socionics. Scalability of complex social systems. Springer, Heidelberg, Germany, pp 84–103 24. Schillo M, Fischer K, Fley B, Florian M, Hillebrandt F, Spresny D (2004) FORM – A sociologically founded framework for designing self-organization of multiagent systems. In: RASTA 2002, Springer, Bologna, Italy, pp 156–175 25. Schuldt A (2009) Decentralisation and interaction efficiency in cooperating autonomous logistics processes. In: LDIC 2009, Springer, Bremen, Germany 26. Schuldt A (2010) Multiagent coordination enabling autonomous logistics. Doctoral dissertation, Universität Bremen, Germany 27. Schuldt A (2011) Team formation for agent cooperation in logistics: protocol design and complexity analysis. In: ICAART 2011, INSTICC Press, Rome, Italy 28. Schuldt A, Werner S (2007) Towards autonomous logistics: conceptual, spatial, and temporal criteria for container cooperation. In: LDIC 2007, Springer, Bremen, Germany, pp 311–319 29. Smith RG (1977) The contract net: a formalism for the control of distributed problem solving. In: IJCAI 1977, William Kaufmann, Cambridge, MA, USA, p 472 30. Timm IJ, Knirsch P, Kreowski HJ, Timm-Giel A (2007) Autonomy in software systems. In: Hülsmann M, Windt K (eds) (2007) Understanding autonomous cooperation and control in logistics. Springer, Heidelberg, Germany, pp 255–273 31. Windt K (2008) Ermittlung des angemessenen Selbststeuerungsgrades in der Logistik – Grenzen der Selbststeuerung. In: Nyhuis P (ed) Beiträge zu einer Theorie der Logistik, Springer, Heidelberg, Germany, pp 349–372 32. Windt K, Hülsmann M (2007) Changing paradigms in logistics. In: Hülsmann M, Windt K (eds) (2007) Understanding autonomous cooperation and control in logistics. Springer, Heidelberg, Germany, pp 1–16 33. Wooldridge M, Jennings NR (1995) Intelligent agents: theory and practice. Knowl Eng Rev 10(2):115–152 34. Wooldridge M, Jennings NR (1999) The cooperative problem solving process. J Logic Comput 9(4):563–592
Part II
Methodical Contributions and Limitations
Chapter 8
Introduction to Methodical Contributions and Limitations of Autonomous Cooperation and Control in Logistics Till Becker and Katja Windt
Autonomous control in logistics has different perspectives. Its application is composed of the technology that enables autonomous control on the one hand and the methods that define how the autonomous units act on the other hand. Both are prerequisite components of autonomous control in logistics. This section focuses on challenges related to autonomous control methods. A narrow view on these methods describes them as “generic algorithms that describe how logistics objects render and execute decisions by their own” [3]. In a broader sense, methodical aspects of autonomous control in logistics involve the selection of the respective autonomous agents within the logistic system, the determination of their goals [1], and the modeling of autonomous control strategy (e.g. in terms of a framework). This also includes the previously mentioned algorithms, the autonomous control methods. Major research challenges pertaining to methodical aspects of autonomous control in logistics can be identified. Comprehensive modeling of autonomously controlled processes is the foundation of further developments in autonomous control [2]. It is important that these models cover all aspects of autonomous control and that they can be embedded in existing modeling approaches in logistics. This is a prerequisite for the integration of autonomous control in today’s logistics applications. Furthermore, the selection of the appropriate autonomous control strategy remains an important question. Although numerous autonomous control methods have been developed in the past, it is still necessary to distinguish between different logistic scenarios when selecting an autonomous control method. Additionally, scalability of autonomously controlled logistic systems is also an issue. The behavior of autonomous control can be subject to changes in a dynamic logistic environment. Therefore it is essential to investigate the influence of fluctuating parameters on the performance of autonomous control. Finally, computational limitations play an important role for the implementation of autonomous control. Although technological advances could offer increased computational capacity,
T. Becker (B) and K. Windt Global Production Logistics, International Logistics, School of Engineering and Science, Jacobs University Bremen, Germany e-mail:
[email protected] M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_8,
93
94
T. Becker and K. Windt
computational power remains a rare resource. Sophisticated methodical approaches on how to utilize information technology in autonomous control in logistics need to be available. As mentioned before, different perspectives need to be considered when applying autonomous control in a logistic system. The presented challenges are mainly situated in the methodical part of autonomous control, but there are also overlaps with organizational and technological issues. However, the relevance of the investigation of challenges and limitations from the methodical perspective is obvious: firstly, it is important to figure out at which point of logistics parameter combination autonomous control is no longer beneficial, either due to lack of performance or due to high implementation costs in relation to its benefits. Secondly, if the methodical limitations are known it is possible to search for solutions in the technical or organizational area. The methodical challenges call for the definition of research goals whose achievement helps to overcome those limitations. The first goal is the development of a modeling methodology that is able to depict autonomous control in logistics. Furthermore, the modeling methodology needs to be compatible to existing modeling approaches. This ensures legibility of the new models for modeling experts and the reusability of existing process models when adapting autonomous control. Another goal is the systematic mapping of autonomous control methods to specific application scenarios, either by developing a catalogue or by assigning certain characteristics of autonomous control methods to characteristics of logistic scenarios. A third goal is the investigation of computational limitations of autonomous control. An often specified advantage of autonomous control is its applicability when centralized planning systems are no longer able to provide a feasible solution in a reasonable amount of time, e.g. in very complex and dynamic scenarios. However, autonomous control uses distributed computational power and is also affected by computational limitations. Autonomous control can only be beneficial when it doesn’t exceed its computational capabilities. All of the following articles address methodical issues of autonomous control in logistics. They aim at getting closer to the previously defined goals. The contribution of Bernd Scholz-Reiter, Michael Görges, and Thomas Jagalski is entitled “Logistic Systems with Multiple Autonomous Control Strategies”. The authors analyze the interaction of multiple autonomous control methods applied in a single logistic scenario. It is pointed out that the combination of autonomous control methods does not necessarily lead to an identical performance when applying the methods individually. The selection of autonomous control methods always calls for a comprehensive view on the whole logistic system. Hans-Jörg Kreowski, Sabine Kuske, Melanie Luderer, and Caroline von Totth focus on opportunities and limitations of computation. Their article “Communities of Autonomous Units – An Approach to Interactive Computation, its Power and Limitations” deals with computational complexity. They consider communities of autonomous units in a formal framework in order to make statements about computational complexity of autonomous control.
8
Introduction to Methodical Contributions and Limitations
95
The article “Potentials and Limitations of Autonomously Controlled Production Systems” by Bernd Scholz-Reiter, Michael Görges, and Henning Rekersbrink investigates the performance of autonomous control methods under varying degrees of system complexity and dynamics. The results are compared with the performance of classical production planning and control approaches. Their research addresses the goal of determining the most appropriate autonomous control method for a specific logistic scenario. Modeling and scalability are the subjects of the contribution “Scalability Effects in Modeling Autonomously Controlled Logistic Processes – Challenges and Solutions in Business Process Modeling” by Bernd Scholz-Reiter, Daniel Rippel, and Steffen Sowade. They present issues that are raised when the modeling methodology has to handle scalability of logistic processes. Furthermore, solutions to the issues are proposed based on two exemplary logistic scenarios of different scale. Nicolas Gebhardt, Oliver Jeken, and Katja Windt address the enablers of autonomous control in logistics in their article “Exploitation of ManufacturingFlexibilities in Decision Methods for Autonomous Control of Production Processes – Findings from Industrial Practice and Theoretical Analysis”. They claim that manufacturing flexibility provides decision alternatives for autonomous control as it offers multiple ways to perform a manufacturing process. Therefore, the available flexibilities determine how autonomous control needs to be applied in a logistics process. They support their findings about the utilization of flexibility with an observation from industrial practice.
References 1. Nyhuis P, Wiendahl H (2008) Fundamentals of production logistics: theory, tools and applications. Springer 2. Scholz-Reiter B, Kolditz J, Hildebrandt T (2009) Engineering autonomously controlled logistic systems. Int J Prod Res 47(6):1449–1468 3. Windt K, Becker T, Jeken O, Gelessus A (2010) A classification pattern for autonomous control methods in logistics. Logisti Res 2(2):109–120
Chapter 9
Logistic Systems with Multiple Autonomous Control Strategies Bernd Scholz-Reiter, Michael Görges, and Thomas Jagalski
9.1 Introduction Production planning and control (PPC) systems have to cope with rising complexity and dynamics that arise from a higher demand for individualized goods, short delivery times, a strict adherence to due dates and internal unexpected events, e.g. machine breakdowns or rush orders. Conventional production planning and control methods cannot handle unpredictable events and disturbances in a satisfactory manner because in practice the complexity of centralized architectures tends to grow rapidly with size, resulting in rapid deterioration of fault tolerance, adaptability and flexibility [8]. One approach to overcome these difficulties is to develop decentralized systems with autonomous control methods to reduce the complexity that has to be taken into account for rendering decisions [13]. Recent developments in information and communication technology, such as radio frequency identification (RFID), wireless communication networks etc., enable intelligent and autonomous logistic objects to communicate with each other and with their resources and to process the acquired information. Combining the autonomous control approach with the developments in information and communication technology may lead to a coalescence of material flow and information flow and enable the logistic objects to manage and control their manufacturing process autonomously [13]. Modeling and benchmarking autonomous control strategies requires dynamic models. Furthermore, one has to consider both, the local decision-making processes as well as the global behavior of the system. The interactions and interdependencies between local and global behavior are called Micro-Macro-Link, which is not trivial to describe and analyze. In a colony of ants for example a single ant has no idea about the whole colony. Its actions are based on a few simple rules. On the other hand, the entire colony consisting of thousands of ants is able to build gigantic B. Scholz-Reiter (B), M. Görges, and T. Jagalski Department of Planning and Control of Production Systems, BIBA, University of Bremen, Germany e-mail:
[email protected] M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_9,
97
98
B. Scholz-Reiter et al.
nests, to find shortest paths between food and nest etc. [9]. This self-organization is a so-called emergent behavior of a complex dynamic system and is not derivable from single characteristics [20]. Previous studies showed the effectiveness of autonomous control for scheduling tasks (e.g. [2, 14–17]) but so far there is no systematical analysis of autonomous machine rules for buffer clearance. This paper addresses the implementation of autonomous control as a service rule within a generic scenario of a flexible flow shop with a pheromone-based scheduling. Its main goal is to show that multiple autonomous control strategies within one logistic system – although it is possible to design and implement them – may lead to non-desired behavior and bad overall logistic performance. To achieve this goal the contribution is structured as follows: Sect. 9.2 offers an overview of autonomous control for scheduling tasks in production logistics. In Sect. 9.3 a generic exemplary scenario of a flexible flow shop, its modeling details as well as the chosen autonomous control strategy, i.e. a pheromone-based scheduling heuristic, are presented. Section 9.4 describes the design of two autonomous service rules and evaluates the simulation results in comparison to FIFO service rule, showing, that multiple autonomous control strategies within one logistic system may lead to a dilemma. Section 9.5 shows how to solve the dilemma with the help of a simple non-autonomous service rule and evaluates on the necessity of strictly local information for autonomous control by presenting a correction term to the pheromone-based scheduling. Section 9.6 summarizes the results and offers an outlook on future research.
9.2 Autonomous Control for Flexible Flow Shop Scheduling The main goal of flexible flow shop scheduling is to sequence and assign a set of jobs to a set of production resources in an optimal manner [1]. However, for most instances of this problem class optimal solutions cannot be found within an appropriate time, because these problems are usually NP complete [7, 12]. Instead of finding optimal solutions, different heuristics, e.g. autonomous control strategies, have been developed to derive acceptable solutions. Autonomous control is defined by: ‘Autonomous Control describes processes of decentralized decision-making in heterarchical structures. It presumes interacting elements in non-deterministic systems, which possess the capability and possibility to render decisions independently. The objective of Autonomous Control is the achievement of increased robustness and positive emergence of the total system due to distributed and flexible coping with dynamics and complexity’ [20]. In the context of engineering science, this global definition is adapted: ‘Autonomous Control in logistic systems is characterized by the ability of logistic objects to process information, to render and to execute decisions on their own’ [20]. Autonomous control aims at increasing robustness and performance of logistic systems [21]. Thus, autonomous control strategies incorporate elements that are able to render decisions by themselves using distributed local information. Consequently, the concept of autonomous control requires on one hand
9
Logistic Systems with Multiple Autonomous Control Strategies
99
logistic objects that are able to receive local information, process this information, and make a decision about their next action – and all that as local as possible. On the other hand, the logistic structure has to provide distributed information about local states and different alternatives to both enable decisions in general and to enable sophisticated decisions that offer an acceptable solution. According to a classification, introduced by Windt and Becker (2009), these local information methods can be grouped as follows: rational strategies, bounded rational strategies and combined strategies [22]. This classification is based on the underlying decision mechanisms used by the different autonomous control methods. Rational strategies utilize rational measures for decision making. Bioanalogous control strategies belong to the group of bounded rational strategies. They aim at transferring fundamental mechanisms of biologic self-organizing systems to autonomous decision making methods. Thus, autonomous control strategies may rely on information about the current situation and a prediction of a future situation of the system (expected values) or on information about how good alternatives had been in the past (experience of the predecessors) or on both. One group of autonomous control strategies that rely on experience of predecessors are bioanalogous control strategies. In literature one can find several attempts to explain the emergent behavior of large scale structures in biological systems. Camazine et al. (2001) offer an overview and some case studies of self-organization in biological systems. The case studies comprise social insects, slime moulds, bacteria, bark beetles, fireflies and fish [5]. According to the authors biological self-organization can be found in group-level behavior that arises in most cases from local individual actions that are influenced by the actions of neighbors or predecessors and in structures that are build conjointly by individuals. Colonies of social insects, e.g. ants or honey bees, show an impressive behavior, which has been classified as Swarm-Intelligence [5]. The individuals follow simple rules that allow solving complex problems beyond the capabilities of single group members. These colonies are characterized by adaptiveness, robustness and self-organization [5]. Several of these rational and biologically inspired autonomous control strategies have been applied to flexible flow shop problems, e.g. the queue length estimator, the pheromone-based control strategy, the honey bee method and mixed strategies. The queue length estimator (QLE) is an autonomous control strategy that enacts a part to compare actual buffer levels of different alternatives (all parallel machines) that are able to perform its next production step [14]. Buffer levels are calculated as the sum of the estimated processing times of the waiting parts in the respective buffer on the respective machine plus its own expected processing time. When a part has to render the decision about its next processing step it compares the current buffer levels, i.e. the estimated waiting time until processing, and chooses the buffer with the shortest waiting time. Thus, the QLE uses the available information to predict the systems future state. The QLE can be used for scenarios with different processing times as well as scenarios with set-up times. The pheromone-based autonomous control strategy [2] utilizes data from past events. Every time a part leaves a machine, i.e. after each processing step, the part leaves information about the duration of its processing and waiting time at the
100
B. Scholz-Reiter et al.
respective machine in form of an artificial pheromone. The following parts use these data to render their decisions. Thus, the parts’ decisions are based on backward propagated information about the throughput times of finished parts for different routes. Routes with shorter throughput times attract parts to use these routes again. This process can be compared to ants leaving pheromones on their way to communicate with following ants. As in other pheromone concepts (e.g. [3, 10]), the communication takes place indirectly by changing the environment. The parts have to be able to access updated information about throughput time only. Thus, this pheromone-based autonomous control strategy differs from approaches from ant colony optimization (e.g. ACO [13]) since there is no self-reinforcing guided search process for optimal solutions. The pheromone concentration depends on the evaporation of the pheromone and on the time previous parts had to spend waiting in the buffer in addition to the processing time on the respective machine as well as the throughput time. Clearly, the fine-tuning of the evaporation constant for the artificial pheromone is crucial. The pheromone-based autonomous control strategy can be used for scenarios with different processing times. However, in a pheromonebased concept, set-up times are somewhat hard to handle because predecessors’ decisions have influence on successors, which is ordinary not communicated by the pheromone. This can be solved by the introduction of a correction term for the pheromone concentration [15]. The honey bee concept has been adapted to flexible flow shop scheduling problems as well [16]. It mimics the food foraging behavior of honey bees, which is slightly different compared to the pheromone concept. Bees that are aware of a food source can advertise the source in order to recruit nest mates by performing a ‘waggle dance’. With the help of the dance, the bee conveys information about the known food source to the ‘onlooking’ bees, i.e. its general direction, distance, and quality [4]. The length of such a dance is proportional to source quality [18]. In a flexible flow shop scheduling scenario a part advertises a good way for following parts after each processing step and the better the alternative is, e.g. the shorter the throughput time was, the longer the advertisement should be. A homecoming collecting bee evaluates the food source by means of the ratio of energy consumption to the energy conveyed to the hive in form of sugar concentration. The better the individual evaluation of the food source quality is the more dance runs the bee will perform [18, 19]. Thus, the more runs the dance has, the longer the advertisement takes and the more unemployed bees can watch it and are attracted to the best food sources. This is different from the pheromone concept because a single ant does no evaluation at all. When transforming the honey bee concept to a flexible flow shop scheduling scenario one would implement this individual evaluation process as well. Thus, the advertising of a good alternative is not decreased by an exponential decay as it is in a pheromone concept but according to the individual evaluation and decision on the number of waggle dances. The different autonomous control strategies can be combined to a mixed strategy that incorporates a weighted combination of the prediction of the future state of the system and the experience of predecessors (e.g. QLE and the pheromone-based autonomous control strategy, [15]).
9
Logistic Systems with Multiple Autonomous Control Strategies
101
So far there is no systematical analysis of autonomous machine rules for buffer clearance. In order to analyze the behavior of a logistic system with multiple autonomous control strategies for scheduling and buffer clearance, a generic scenario of a flexible flow shop with set-up times was established and the pheromonebased autonomous scheduling method was chosen.
9.3 Exemplary Scenario – Modeling Details The considered exemplary scenario is a matrix-like flow-line manufacturing system producing k different product types at the same time. Each of the products has to undergo m production stages. For each of these production stages there are n parallel production lines available. Therefore, the shop floor consists of mn machines. The raw materials for each product enter the system via sources and the final products leave the system via drains. The production lines are coupled at every stage and every line is able to process every type of product within a certain stage. Switching product types requires a set-up. At each production stage a part has to make an autonomous decision to which of the lines to go to in the next stage. Each machine has an input buffer in front, containing items of the k product types as Fig. 9.1 shows [14]. This scenario was chosen because of its generic and universal character, it can be applied to the majority of real world flexible flow shop configurations.
Source
B11
B12
M11
M12
…
Drain
B1n
Possible next processing step
M1n
B Buffer B22
B2n
M21
M22
M2n
Bm1
Bm2
Bmn
Mm1
Mm2
Mmn
M Machine
…
B21
Fig. 9.1 Generic m n shop floor scenario [14]
102
B. Scholz-Reiter et al.
Table 9.1 Set-up times of the 3 3 machines model Set-up times Machine [min] Mm1 Mm2 Mm3 A!B 30 10 60 A!C 60 30 10 B!A 10 60 30 B!C 60 30 10 C!A 10 60 30 C!B 30 10 60
To handle the complexity, the simulation model is reduced to 3 3 machines producing 3 different product types A; B and C . The model is build with Vensim DSS computer simulation software. The arrival functions for the three product types are defined as sine functions as a representation of the seasonal varying market demand. They are identical except for a phase shift of 1/3 period for the three product types. To model a usual workload of about 80% in real production systems, a mean arrival rate of 0:4 1= h and an amplitude of the sine functions of 0:15 1= h are chosen. It is assumed that the processing times for each product are the same: 120 min. Table 9.1 shows the set-up times for the three parallel machine types 1, 2 and 3 at all production stages m and the three product types respectively. To analyze the logistics performance of the global system, the aggregate buffer level for the three different product types of the first production step was chosen as the logistic performance indicator. As there are different product types with different set-up times, the machines’ service rule for the different product types is important. For the first simulation scenario it is first in – first out (FIFO). To analyze the behavior of a flexible flow shop with more than one autonomous control strategy, the pheromone-based autonomous control strategy was chosen for scheduling. It uses data from past events in a way that every time a part leaves a machine after being processed, it leaves information about the duration of its processing and waiting time at the respective machine as an artificial pheromone. The following parts can use this information to render their decisions. Thus, the parts’ decisions are based on backward propagated information about the throughput times of finished parts for different routes. Routes with shorter throughput times attract parts to use these routes again. To mimic the behavior of ants to search for shortest ways with the help of a random walk, the parts deviate from the decision to simply follow the strongest pheromone concentration with a certain probability (here 5%). The pheromone concentration update algorithm works as follows: Let Pmnk .t/ denote the pheromone concentration for machine mn at time t; Emnk the evaporation constant .0 < Emnk 1/ for product type k at machine mn, ˇmnk a (constant) gain for the pheromone concentration update for product type k at machine mn and TPT mnk .t/ the actual throughput time for product type k at machine mn. Then the pheromone updating process is given by:
9
Logistic Systems with Multiple Autonomous Control Strategies
103
Pmnk .t/ D Pmnk .t/ Pmnk .t 1/Emnk ˇ TPT mnk .t/; if ‘machine has completed its job’ D true C mnk 0; else This flexible flow shop scenario with autonomous scheduling was implemented with the help of the continuous System Dynamics Vensim DSS computer simulation software. The term continuous denotes the continuous material flow and evaporation process, which differs from the flow of discrete parts in e.g. a discrete event simulation model; the simulation time is discrete. In literature, continuous flow models of production systems are often called hybrid models (e.g. [6, 11]). That means the material flow is modeled as continuous flow which is controlled by discrete actions. This discrete control is typical for production systems. The implementation of the different autonomous service rules together with the simulation results are described in the following section.
9.4 Design of Autonomous Service Rules and Simulation Results The simulation model is designed in a way that it allows the analysis of different service rules. First, and for comparison, the flexible flow shop with pheromonebased autonomous control is combined with FIFO service rule in Sect. 9.4.1. In Sect. 9.4.2 a pheromone-based service rule is presented, followed by a QLE service rule in Sect. 9.4.3. An evaluation of the results is given in Sect. 9.4.4.
9.4.1 Simulation Results with FIFO Figure 9.2 shows the aggregate buffer levels of the first production step of the flexible flow shop with pheromone-based autonomous control for scheduling and FIFO as the service rule. The maximum inventory is 13.26 pieces and the mean inventory is 8.65 pieces with a standard deviation of 6.11 pieces.
9.4.2 Simulation Results with a Pheromone-Based Autonomous Service Rule To improve the logistic performance the service rule should be altered. The approach to implement an autonomous service rule, i.e. let the machines select the parts and organize the set-ups, is promising, because set-ups are random in a FIFO scenario. The pheromone-based autonomous service rule works as follows: Setup to the product type with the highest pheromone concentration and with some
104
B. Scholz-Reiter et al.
Pheromone-based scheduling with FIFO service rule 20
10
0
0
30 Simulation Time [days]
Parts in Buffer Average
Fig. 9.2 Aggregate buffer levels of the first production step with FIFO
Pheromone-based scheduling and pheromone service rule
20
10
0
0
30 Simulation Time [days]
Parts in Buffer Average
Fig. 9.3 Aggregate buffer levels of the first production step with pheromone-based service rule
probability (here 5%) set-up to a random product type. The pheromone update process is straightforward: Whenever a part has been processed it leaves an artificial pheromone according to its product type at the machine, which is constantly evaporating. Figure 9.3 depicts the aggregate buffer levels of the first production step of the flexible flow shop with pheromone-based autonomous control for scheduling and the pheromone-based service rule.
9
Logistic Systems with Multiple Autonomous Control Strategies
105
The maximum inventory is 14.74 pieces and the mean inventory is 10.37 pieces with a standard deviation of 8.41 pieces. This performance is not satisfying; even the FIFO service rule shows a better performance within this scenario.
9.4.3 Simulation Results with QLE as Autonomous Service Rule To further analyze the behavior of the flexible flow shop with multiple autonomous control strategies, the QLE is implemented as service rule. The QLE service rule is designed in a way that the machine calculates the total processing times, waiting times plus set-up times of each product type in its buffer. Then, it compares the values and chooses to set-up to the product type with the longest overall processing, waiting and set-up time to maximize periods without set-ups. The aggregate buffer levels of the first production step of the flexible flow shop with pheromone-based autonomous control for scheduling and QLE service rule are shown in Fig. 9.4. The maximum inventory is for pheromone-based scheduling and QLE service rule is 13.48 pieces and the mean inventory is 9.02 pieces with a standard deviation of 8.33 pieces. This performance is not good as well. Although better than the pheromone service rule, the QLE performance is slightly worse than FIFO.
9.4.4 Evaluation of the Results At a first glance, it seems to be surprising that the performance of the autonomous service rules is bad and even below FIFO. When analyzing the drawbacks of the two new autonomous service rules, their lack of performance can easily be explained.
Pheromone-based scheduling with QLE service rule 20
10
0
0
30 Simulation Time [days]
Parts in Buffer Average
Fig. 9.4 Aggregate buffer levels of the first production step with QLE service rule
106
B. Scholz-Reiter et al.
The pheromone-based service rule handles set-ups according to the predecessors experience and does not set-up to actual needs. For example, a set-up would be from product type A to product type B if there are no parts of product type A in the buffer and set-ups to product type B have been a good decision in the past. This decision is a very bad one if there are no parts of product type B in the buffer. A second drawback of the pheromone-based service rule is the deviation from the strongest pheromone concentration (analogous to the random walk of ants), which leads to completely senseless set-ups. These two drawbacks end up in the bad performance shown in Sect. 9.4.2. The pheromone-based autonomous control for scheduling does not take the pheromone-based service rule’s drawbacks into account as there is no synchronization between the different autonomous control strategies. The QLE service rule has two drawbacks as well: First, newly arriving parts lets the QLE re-calculate the overall time to clear the buffer from the respective product type. This leads to many set-ups according to the re-evaluated simulation. Another drawback of the QLE service rule is that the third-best alternatives, i.e. the ones with the longest set-up time, are processed in a subordinate way compared to the other service rules, because differing set-up times are not regarded at all in the scenario with pheromone-based service rule or FIFO. The pheromone-based autonomous control for scheduling does not take the QLE service rule’s drawbacks into account either as there is again no synchronization between the different autonomous control strategies here. Both logistic systems with a pheromone-based autonomous scheduling strategy and autonomous service rules show a very bad performance. A designer of autonomous control strategies is in a dilemma: Multiple autonomous control strategies within one logistic system do not perform well per se, neither do they synchronize themselves and without paying attention, the overall performance can be bad.
9.5 Solution to the Dilemma Exemplarily, two different solutions to the dilemma described in Sect. 9.4.4 are presented in the following. First, an easy and non-autonomous service rule is implemented, which shows a very good performance. Another improvement can be achieved by introducing a correction term to the scheduling pheromone in order to synchronize the autonomous control strategies for scheduling and the autonomous control strategy for buffer clearance.
9.5.1 A Simple But Good Service Rule A simple non-autonomous service rule is implemented. It minimizes set-ups in the following way: As long as there are parts of the type the machine is set-up to: do not make a set-up, go on processing. If a set-up is needed, switch to the type which is quantitatively best represented in the buffer. Figure 9.5 shows the aggregate buffer
9
Logistic Systems with Multiple Autonomous Control Strategies
107
Pheromone-based scheduling and simple service rule 20
10
0 0
30 Simulation Time [days]
Parts in Buffer Average
Fig. 9.5 Aggregate buffer levels of the first production step – pheromone-based scheduling with correction term and simple service rule
levels of the first production step of the flexible flow shop with pheromone-based autonomous control for scheduling and the simple service rule. The maximum buffer level is reduced to 8.83 pieces. The mean buffer level is 5.77 pieces with a standard deviation of 3.88 pieces. The implementation of the simple service rule to let the machines select parts according to the mentioned scheme leads to an improved performance.
9.5.2 Introducing a Pheromone Correction Term Because of two reasons this performance seems to be improvable: The pheromone concentration does not include information about the set-up status of the machine and, a part’s decision can be both, good or bad, depending on how many set-ups the machine has to perform before the part can be processed. The second reason is not included in the pheromone concentration either. Thus, the machines’ service rule has to be improved and a correction term for the pheromone concentration has to be implemented [15]. A correction term is introduced to the update process of the pheromone concentration. This correction term includes information about the product type a machine is set-up to after a part has been processed. This can not be done by simply leaving a higher amount of the pheromone because this additional information should effect a direct successor’s decision only. A higher pheromone quantity would evaporate over time according to the evaporation constant leading to bad information for the next but ones’ decisions. Thus, the correction term consists of an increasing of the pheromone concentration but with a higher evaporation constant. The pheromone update algorithm works as follows: Let CT mnk .t/ denote the value of
108
B. Scholz-Reiter et al.
the correction term for product type k at machine mn at time t; ımnk a constant adjusted to the execution time for product type k at machine mn, ECmnk the evaporation constant for the correction term .1 > EC E/ for product type k at machine mn and set-up_statusmn.t/ the status the machine mn is actually set-up to. Then, the pheromone concentration with correction term P_cormnk .t/ consists of the pheromone part P_partmnk .t/ and the correction term part C Tmnk .t/: P _cormnk .t/ D P _ partmnk .t/ C CT mnk .t/ with P _ partmnk .t/ D P _ partmnk .t/ P _ partmnk .t 1/Emnk ˇ TPT mnk .t/; if ‘machine has completed its job’ D true C mnk 0; else and CT mnk .t/ D CT mnk .t/ CT mnk .t 1/ECmnk ı ; if set-up_statusmnk .t/ D k C mnk 0; else The (higher) evaporation constant for the correction term ECmnk is adjusted to the execution time (processing time plus set-up time) of the next part on a particular machine in order to improve the overall performance of the logistic system. Figure 9.6 shows the aggregate buffer levels of the first production step with
Pheromone with correction term and simple service rule 20
10
0 0
30 Simulation Time [days]
Parts in Buffer Average
Fig. 9.6 Aggregate buffer levels of the first production step with pheromone-based scheduling with correction term and simple service rule
9
Logistic Systems with Multiple Autonomous Control Strategies
109
Table 9.2 Performance measures of the different service rules and the altered pheromone-based scheduling strategy FIFO Pheromone based QLE service Simple non-autonomous Pheromone with (4.1) service rule (4.2) rule (4.3) service rule (5.1) correction term (5.2) Mean 8:65 10:37 9:02 5:77 8:55 Max 13:26 14:74 13:48 8:83 5:51 STD 6:11 8:41 8:33 3:88 3:67
pheromone-based scheduling with correction term and the simple new service rule for buffer clearance 5.1). With the help of the pheromone correction term the maximum buffer level is reduced to 8.55 pieces. The mean buffer level goes down to 5.51 pieces with a standard deviation of 3.67 pieces. The implementation of the pheromone correction term pays as the comparison of FIFO (cf. 4.1), pheromone-based service rule (cf. 4.2), QLE service rule (cf. 4.3) and simple service rule (cf. 5.1) in Table 9.2 summarizes. One has to keep in mind that introducing a correction term to the scheduling pheromone means abandoning the use of strictly local data because of the interaction between the machines (their set-up status is local information) and the pheromonebased autonomous control strategy for scheduling.
9.6 Summary and Outlook A generic matrix model of a flexible flow shop with set-ups and a pheromonebased autonomous control strategy was presented to analyze the performance and behavior of multiple autonomous control methods within one logistic system. Two autonomous service rules for buffer clearance, a pheromone-based service rule and the QLE service rule were introduced and implemented into a System Dynamics computer simulation model. The simulation results were compared to the FIFO service rule and it was shown that the overall performance of the autonomous service rules was bad, even worse than FIFO, in this scenario. One big result of this contribution is that the application of autonomous control has its limitations: Multiple autonomous control strategies within one logistic system do not perform well per se, neither do they synchronize themselves. Designers of autonomous control strategies should beware: Autonomous control should not be implemented for the sake of autonomous control itself. Designing autonomous control methods according to current requirements as well as synchronization is highly needed (and possible). Two solutions to the dilemma were introduced: First, a simple service rule for buffer clearance showed a better performance. Second, the interaction between buffer clearance and scheduling was improved by introducing a pheromone correction term to the autonomous control strategy for scheduling. This course of action showed on the one hand an improved performance, on the other hand leads this
110
B. Scholz-Reiter et al.
manual synchronization of the different autonomous control strategies to a loss of the paradigm to strictly use local data. In order to use the pheromone correction term, the local information about the machines’ set-up status must be propagated to the pheromone-based autonomous control strategy for scheduling. Of course, not all autonomous service rules would have a bad performance in a logistic system with multiple autonomous control strategies. Even a pheromonebased service rule without random walk could probably perform in a better way. This shows that future research is needed: Better autonomous service rules should be designed and systematically analyzed. Additionally, a further achievement in performance could be made by mixed autonomous control strategies for scheduling or as service rules. Furthermore, machine breakdowns or rush orders should be analyzed in order to analyze the robustness against internal dynamics of logistic systems with multiple autonomous control strategies.
References 1. Allahverdi A, Ng CT, Cheng TCE, Kovalyov M (2008) A survey of scheduling problems with setup times or costs. Eur J Oper Res 187(3):985–1032 2. Armbruster D, de Beer C, Freitag M, Jagalski T, Ringhofer C (2006) Autonomous control of production networks using a pheromone approach. Phys A 363(1):104–114 3. Bonabeau E, Dorigo M, Theraulaz G (1999) Swarm intelligence – from natural to artificial Systems. Oxford Press 4. Camazine S, Sneyd J (1991) A model of collective nectar source selection by honey bees: self-organization through simple rules. J Theor Biol 149:547–571 5. Camazine S, Deneubourg JL, Franks NR, Sneyd J, Theraulaz G, Bonabeau E (2001) Selforganization in biological systems. Princeton University Press, New Jersey 6. Chase C, Serrano J, Ramadge P (1993) Periodicity and chaos from switched flow systems. IEEE Trans Automat Contr 70–83 7. Garey MR, Johnson DS (1979) Computers and intractability: A guide to the theory of NPcompleteness. Freeman, San Francisco 8. Kim JH, Duffie NA (2004) Backlog control for a closed loop PPC system. Ann CIRP 53(1):357–360 9. Parunak van Dyke H (1997) Go to the ant. Ann Oper Res 75:69–101 10. Peeters PV, Brussel H, Valckenaers P, Wyns J, Bongaerts L, Heikkilä T, Kollingbaum M (1999) Pheromone based emergent shop floor control system for flexible flow shops. In: Proceedings of the International Workshop on Emergent Synthesis, pp. 173–182 11. Peters K, Worbs J, Parlitz U, Wiendahl H-P (2004) Manufacturing systems with restricted buffer sizes. Nonlinear dynamics of production systems. Wiley, Weinheim 12. Pinedo ML (2008). Scheduling – theory, algorithms, and systems. Springer, New York 13. Scholz-Reiter B, Windt K, Freitag M (2004) Autonomous logistic processes. New Demands and First Approaches. Proceedings of the 37th CIRP International Seminar on Manufacturing Systems, pp. 357–362 14. Scholz-Reiter B, Freitag M, de Beer C, Jagalski T (2005) Modelling dynamics of autonomous logistic processes: Discrete-event versus continuous approaches. CIRP Ann Manuf Technol 55(1):413–416 15. Scholz-Reiter B, Jagalski T, de Beer C, Freitag F (2007) Autonomous shop floor control considering set-up times. In: Proceedings of the 40th CIRP International Seminar on Manufacturing Systems. Liverpool, Great Britain
9
Logistic Systems with Multiple Autonomous Control Strategies
111
16. Scholz-Reiter B, Jagalski T, Bendul J C (2008) Autonomous control of a shop floor based on bee’s foraging behaviour. In: Haasis HD, Kreowski HJ, Scholz-Reiter B (eds.) Dynamics in logistics, pp 415–423. Berlin Heidelberg, Germany 17. Scholz-Reiter B, Görges M, Jagalski T, Naujok L (2010) Modelling and analysis of an autonomous control method based on bacterial chemotaxis. In: Proceedings of the 43rd CIRP International Conference on Manufacturing Systems. Neuer Wissenschaftlicher Verlag, Wien, pp 699–706 18. Seeley T, Camazine S, Sneyd J (1991) Collective decision-making in honey bees: how colonies choose among nectar sources. Behavioral Ecology Sociobiology 28:277–290 19. Seeley TD (1994) Honey bee foragers as sensory units of their colonies. Behavioral Ecology Sociobiology 34:51–62 20. Windt K, Hülsmann M (2007) Changing paradigms in logistics – understanding the shift from conventional control to autonomous cooperation and control. In: Hülsmann M, Windt K (eds.) Understanding autonomous cooperation & control – the impact of autonomy on management, information, communication, and material flow, pp 4–16. Springer, Berlin 21. Windt K, Böse F, Philipp T (2008) Autonomy in production logistics – identification, characterisation and application. Int J Robot CIM 24(4):572–578 22. Windt K, Becker T (2009) Applying autonomous control methods in different logistic processes – a comparison by using an autonomous control application matrix, Proceedings of the 17th Mediterranean Conference on Control and Automation. Thessaloniki, Greece
Chapter 10
Communities of Autonomous Units: An Approach to Interactive Computation, Its Power and Limitations Hans-Jörg Kreowski, Sabine Kuske, Melanie Luderer, and Caroline von Totth
10.1 Introduction In the world of logistics, one must deal with quite a variety of aspects including the interplay of production and transportation, the interoperation of trucks, trains, vessels and planes, the wide spectrum of used technologies like radio frequency identification, telematics, global positioning, and networking. Moreover most logistic affairs concern various parties with different and partly conflicting interests. Logistic networks and systems become more and more complex and dynamic and they often consist of many interacting processes without central control. Therefore, the modeling of logistic networks and systems requires concepts that allow one to describe interaction, cooperation, communication, and coordination of logistic processes with a high degree of autonomy. The graph- and rule-based framework of communities of autonomous units generates modeling devices to meet these objectives (e.g., [9, 11, 13]). In this sense, it is an attempt to realize the paradigm of cooperating autonomous entities like the approaches of multi-agent systems (e.g., [1, 6, 7, 17, 18]). A community comprises a set of autonomous units that act and interact in a common environment. An autonomous unit specifies an autonomous process. It consists of a set of rules that specify the potential actions and a control condition based on which a unit decides about its next action. Moreover, a unit has a goal that it tries to reach. The common dynamic environments are described by graphs that are changed by rule applications while the units run their processes interleaving the various actions sequentially or performing them in parallel or concurrently. In addition to its units, a community provides the possible initial environments and may also have an overall goal to be reached by the cooperation of the units. Finally, there may be a global control condition that can coordinate the communication and cooperation among the units. Clearly, this can restrict the autonomy of the single units.
H.-J. Kreowski (B), S. Kuske, M. Luderer, and C. von Totth Department of Computer Science, University of Bremen, Germany e-mail:
[email protected] M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_10,
113
114
H.-J. Kreowski et al.
But such limitations are acceptable as autonomy is not meant to force a chaotic behavior. The main aim of communities of autonomous units is the modeling of the interaction and cooperation of logistic processes. Examples of this kind are production and transportation in production networks, communication-based dynamic scheduling, pickup and delivery vehicle routing and capacitated vehicle routing based on ant colony optimization [5, 10, 12, 14]. But a community can also be interpreted as a computational model if all its features are algorithmic including the rule applications, the decisions made due to the control conditions, the construction of initial environments and the tests whether a goal is reached. An environment that satisfies the overall goal may be considered as terminal so that one can interpret the runs from initial to terminal environments as computations and the corresponding relation between initial and terminal environments as the computed relation. To find out about the computational power, Turing machines can be transformed into communities of autonomous units so that the latter framework inherits the computational completeness from the former one. A similar result is obtained with respect to the famous complexity class NP that contains all decision problems that can be solved nondeterministically in polynomial time. The class NP is of great importance for the area of logistics because most logistic planning and optimization problems are closely related to NP problems. As all NP problems can be solved by polynomial Turing machines, their transformation yields polynomial communities of autonomous units which, in turn, cover the class NP fully. These computational aspects are further studied in this paper. It turns out that communities of autonomous units do not only allow the formal modeling of interactive logistic processes, but also their computational analysis. The paper is organized as follows. In Sect. 10.2, a short survey is given on communities of autonomous units. In Sect. 10.3, the concept is illustrated by modeling a typical scheduling problem of the job-shop kind. Section 10.4 is devoted to the computational aspects of communities of autonomous units. The paper ends with some concluding remarks.
10.2 Autonomous Units Autonomous units interact in a common environment which is modeled as a graph. As a basic modeling device, an autonomous unit consists of a set of graph transformation rules, a control condition, and a goal. The graph transformation rules specify all actions the unit can perform. Such an action can be a movement of the autonomous unit within the current environment, the exchange of information with other units via the environment, or local changes of the environment. The control condition regulates the application process. For example, it may require that a sequence of rules be applied as long as possible. The goal of a unit determines how the transformed graphs should look like eventually.
10
Communities of Autonomous Units
115
10.2.1 Environments The environments where a community of autonomous units operates belong to an underlying graph class G that can be chosen freely and that may differ according to the problem to be modeled. Concretely, the class G may comprise directed or undirected, labeled or unlabeled, cyclic or acyclic, connected, complete or discrete graphs, trees, Petri nets, etc. A concrete example of a graph class G is the set of all edge-labeled directed graphs defined as follows. Graphs: 1. A (directed edge-labeled) graph is a system G D .V; E; s; t; l/ where V is a set of nodes, E is a set of edges, the mappings s; t W E ! V assign to every edge its source s.e/ and its target t.e/; and the mapping l assigns a label to every edge in E. 2. The components of G are also denoted by VG , EG , sG , tG , and lG . 3. An edge is called a loop if its source and target coincide. 4. A graph H is a subgraph of G, denoted by H G if VH VG ; EH EG and sH ; tH , and lH are the restrictions of sG ; tG , and lG to EH . 5. A graph morphism g W L ! G from a graph L to a graph G consists of two mappings gV W VL ! VG ; gE W EL ! EG such that sources, targets and labels are preserved, i.e. for all e 2 EL ; gV .sL .e// D sG .gE .e//; gV .tL .e// D tG .gE .e//; and lL .e/ D lG .gE .e//: In the following we omit the subscript V or E of g if it can be derived from the context. In order to obtain well readable graphical representations of environments, nodes may carry a label inside that is technically modeled as a loop. This means that an x-labeled node is modeled as a node with an x-labeled loop. Such nodes are called x-nodes in the following. An example of an environment is given in Fig. 10.1. It represents a set M D fm1 ; : : : ; m4 g of machines and a set J D fj1 ; : : : ; j5 g of jobs so that every job is associated with a machine. For each machine m there is at most one job that is
Fig. 10.1 Example of an environment
116
H.-J. Kreowski et al.
being processed; all other jobs associated with m are waiting. Concretely, the graph in Fig. 10.1 contains an m-node for every machine m and a j -node for every job j . Every job node is connected to exactly one machine node via an edge which is labeled with the residual time the machine needs for accomplishing the job. Every job node has an additional loop which is labeled either with waiting or with selected where the former expresses that the job is waiting for being selected and the latter means that the job is being processed by the associated machine. For example, machine m4 is processing job j3 while job j2 is waiting at m4 . At machine m3 no job is being processed.
10.2.2 Graph Class Expressions and Goals The work of every community starts with an initial environment. For specifying all possible initial environments of a community in a finite way, graph class expressions from an underlying class X are used. A graph class expression is any syntactic device that specifies a set of graphs. For example, every finite subset S G is a graph class expression that specifies itself, i.e. SEM.S / D S . Moreover, every set L of labels specifies the class of all graphs in G the labels of which are elements of L. The least restrictive graph class expression is the term all specifying the class G. Some further graph class expressions used in the examples are explained directly there. Graph class expressions can also be used to specify the goal of an autonomous unit or of a community, i.e., a special situation that should be reached. For explicit use below the goal env.M / D fm1 ; : : : ; mk g specifies the graph with node set fv1 ; : : : ; vk g where every node vi is an mi -node. This graph represents the set M (of machines). Moreover, we use the expression reduced as a goal of an autonomous unit. It specifies all graphs to which no rule of the unit is applicable.
10.2.3 Basic Actions The actions of autonomous units are modeled as applications of graph transformation rules taken from an underlying rule class R that specifies transformations on the graphs in G. There are various types of graph transformation rules such as double-pushout rules, single-pushout rules, edge- or node-replacement rules, relabeling rules, etc. [16]. As a concrete example we consider double-pushout rules with and without negative application conditions as rule class R [8] for a more general form of negative application conditions). Rules: 1. A (graph transformation) rule is a tuple r D .L; K; R/ of graphs such that L K R: L is the left-hand side, K is the gluing graph, and R is the right-hand side of r.
10
Communities of Autonomous Units
117
2. The application of r to a graph G yields a graph G 0 if one proceeds according to the following steps: – Choose a graph morphism g W L ! G so that (1) for all items x; y (nodes or edges) of L with x ¤ y, g.x/ D g.y/ implies that x and y are in K, and (2) the deletion of all items of g.L/ g.K/ does not produce dangling edges. – Delete all items of g.L/ g.K/. The resulting graph is denoted by D. – Add R disjointly to the graph D. – Glue D and R by identifying the nodes and edges of K in R with their images under g. 3. A rule with a negative application condition is a tuple r D .N; L; K; R/ of graphs such that N L K R. The application of such a rule to a graph G is defined as above with the additional condition that g cannot be extended to some morphism g0 W N ! G of which g is the restriction to L. 4. The transformation of G to G 0 via the application of r is denoted by G ) G 0 . A r
sequence of rule applications starting in G and ending in G 0 where each applied
rule stems from some rule set P is called a derivation and is denoted by G ) G 0 . P
The two conditions in point 2 are also called gluing condition. The condition in point 3 assures that the negative context specified by N L is not present in the host graph G. Graph transformation rules without a negative application condition are displayed simply in the form L ! R. Nodes or edges with identical relative positions in L and R belong to the gluing graph K. An example of such a rule is given in Fig. 10.2. Its left-hand side shows a machine m with a selected associated job j that still needs rpt time units on m (rpt stands for residual processing time). The righthand side shows the same graph but where the residual processing time is decreased by 1. The condition rpt > 1 below the arrow indicates that the rule can only be applied if the value of rpt in the matched part of the host graph is greater than 1. (This means that before applying the rule a value greater than 1 must be assigned to rpt.) Graphical representations of rules with negative application conditions are depicted in the form N L ! R. For example, the rule in Fig. 10.3 models the
Fig. 10.2 The rule for processing a job
118
H.-J. Kreowski et al.
Fig. 10.3 The rule for selecting a job
selection of a waiting job j by a machine m provided that there is no other selected job j 0 connected to the same machine.
10.2.4 Parallel Actions Autonomous units may act sequentially, in parallel or concurrently and for each case there are meaningful applications [11, 13]. In the following, we concentrate on parallel actions that are well suited to model logistic problems (e.g., [5, 14]). The basic idea of parallelism in a rule-based framework is the application of many rules simultaneously and also the multiple application of a single rule. To obtain these possibilities, we assume that multisets of rules can be applied rather than single rules. Therefore, an application of a multiset of rules to a graph yields a graph which models the parallel application of the rules in the multiset. For the double-pushout approach this is achieved by disjointly composing the components of the rules in the multiset and applying the resulting rule as described above. Parallel rules: Let r1 ; : : : ; rn 2 R with ri D .Li ; Ki ; Ri /. Then the parallel rule r1 C C rn is defined by .L1 C C Ln ; K1 C Kn ; R1 C C Rn / where C denotes the disjoint union of graphs. This includes the case n D 0 yielding the empty rule .;; ;; ;/, which is always applicable without effect. Parallel rules are rules in the ordinary sense so that their application is already defined. Without going into the technical details, it is not difficult to see that each component rule can be applied to some host graph whenever the parallel rule can be applied. Therefore, negative application conditions can be used for the component rules as before, and we can assume a parallel rule may be composed of rules with and without negative application conditions (see [4] and [8] for more details).
10.2.5 Control Conditions In many cases, rule application is highly nondeterministic – a property that is not always desirable. In order to reduce the degree of nondeterminism of rule
10
Communities of Autonomous Units
119
application, autonomous units can regulate their transformation processes by a control condition taken from an underlying class C. A control condition is any expression that specifies a set of graph sequences. The semantics of control conditions for autonomous units is defined in [14]. The most general control condition is the term free that allows to apply any rule at any time. More restrictive and often used control conditions are regular expressions over rules. Every such regular expression specifies a set of words each of which corresponds to a sequence of rules. This control condition requires to apply the rules of the unit in the same order as they occur in some sequence specified by the control condition. For example, the regular expression r1 I r2 requires to apply first the rule r1 once and afterwards the rule r2 an arbitrary number of times. Another useful control condition is r! which requires to apply rule r as long as possible. We also use the !-operator in combination with regular expressions. For example, r1 I r2 ! demands to apply rule r1 once and then rule r2 as long as possible. Every control condition may have an additional flag busy that restricts the set of allowed derivations to those in which every rule is applied as early as possible. An autonomous unit whose control condition has the busy-flag set is called a busy autonomous unit. In order to control the transformation process of a set of autonomous units, global control conditions are used [14]. Such a control condition may require that a unit must wait until another unit has finished its work or that several units should act in parallel. All control conditions used in the examples are explicitly explained there.
10.2.6 Communities of Autonomous Units A community is a set of autonomous units together with an initial graph class expression, a global control condition and an overall goal. The initial graph class expression specifies the possible initial environments where the autonomous units can start their transformations, the global control condition regulates the interaction of the autonomous units and the overall goal is a graph class expression that specifies which environment should be reached. Hence, a community is a system COM D .Env; Aut; control; goal/, where Env; goal 2 X are graph class expressions called the initial environment specification and the overall goal, respectively, Aut is a set of autonomous units, and control is a global control condition. Every autonomous unit consists of a set of rules, a control condition and a goal. i.e., an autonomous unit is a system aut D .P; c; g/ where P R is a set of rules, c 2 C is a control condition, and g 2 X is the goal. The parallel semantics of a community of autonomous units consists of a set of derivations each starting with an initial environment and applying in every step a parallel rule composed of the rules occurring in the units of the community such that the control conditions of all units as well as the global control condition are satisfied. Every such derivation is called a run of the community. A run is called successful if
120
H.-J. Kreowski et al.
its last environment satisfies the goals of the autonomous units as well as the global goal of the community. A formal parallel semantics is presented in [11].
10.3 Example: Modeling the Job-Shop Scheduling Problem In this section, we model the job-shop scheduling problem as a community of autonomous units in order to illustrate how communities can be used to simulate solution constructions of logistic problems. The job-shop scheduling problem is a classical scheduling problem that consists of a finite set of jobs that must be processed on a finite set of machines (cf., e.g., [2, 3, 15]). Every job is a sequence of machines prescribing the order in which the machines have to be visited by the job for processing. Additionally, there is a processing time for every machine occurring in a job. Formally, the job-shop scheduling problem is a triple JSP D .M; J; pt/ where M is a finite set of machines, J M is a finite set of jobs such that for each j 2 J and each m 2 M; count.m; j / 1,1 and pt W T ! N where T D f.m; j / 2 M J j count.m; j / D 1g is the set of tasks of JSP . The mapping pt associates a processing time with each task. A feasible schedule of JSP is a mapping start W T ! N that assigns a start time to every task in T such that the following holds: 1. The operations of each job j are processed sequentially in the order in which they occur in j , i.e., for each job j D m1 : : : ml 2 J with l 2; start.mi ; j / C pt.mi ; j / start.mi C1 ; j /, for i D 1; : : : ; l 1. 2. No machine can process more than one job simultaneously, i.e., for all tasks .m; j /; .m0 ; j 0 / 2 T with m D m0 and j ¤ j 0 ; start.m; j / C pt.m; j / start.m0 ; j 0 / or start.m0 ; j 0 / C pt.m0 ; j 0 / start.m; j /. The makespan of each schedule start is defined as makespan.start/ D maxfstart.t/ C pt.t/ j t 2 T g which is the maximum finishing time of all tasks. A best schedule is one where makespan is minimized. In this modeling example we do not aspire to construct an optimal solution but to illustrate how the autonomous and interactive behavior of machine and job units represents the construction of feasible schedules. For reasons of space limitations, we do the modeling without an explicit construction of graphs representing the schedules. For the same reasons, we assume that every job consists of at least one machine
1 For alphabets A and B, the mapping count W AB ! N counts how often a symbol of A occurs in a word of B , i.e., for machine m and job j , count .m; j / counts the number of occurrences of m in j .
10
Communities of Autonomous Units
121
and that pt.t/ 2 for each task t. However, it should be noted that these restrictions can be easily dropped by extending the rules of the units accordingly. The community of autonomous units Com.JSP/ in Fig. 10.4 for the job-shop scheduling problem JSP D .M; J; pt/ with the task set T consists of the following components. The initial environment specification init.M; J; pt/ specifies the graph consisting of an m-node for every machine m 2 M , a j -node for every job j 2 J such that every j -node has a waiting-loop and an edge to an m-node if m is the first machine in j . The edge from j to m is labeled with pt.m; j /. Hence, in the beginning, every job is waiting at its first machine for being selected. The set of autonomous units in Com.JSP/ contains a unit macm for every machine m in M and a unit jobj for every job j in J . The global control condition does not restrict anything and the overall goal is equal to e nv.M / which specifies the graph consisting of an m-node for every machine m 2 M . This goal assures that all jobs must be finished at the end of a successful run. For each machine m 2 M , the autonomous unit macm consists of the rules select.m; j / and process.m; j / for every task .m; j /, where the former rule is given in Fig. 10.3 and the latter in Fig. 10.2. As mentioned before, the rule process.m; j / models the processing of j by m (up to the last step) by decreasing the residual processing time by one in every application of process. The rule select.m; j / models the selection of j by m provided that no other job is selected. It also reduces the processing time at the edge from j to m by 1. The control condition of macm requires that after every application of a select-rule the process-rules are applied as long as possible, while select may be applied arbitrarily often. The goal of the machine unit specifies all graphs to which neither select nor process is applicable; this is denoted by the term reduced. The autonomous unit macm is shown in Fig. 10.5.
Fig. 10.4 The community of autonomous units Com(JSP)
Fig. 10.5 The autonomous unit macm
122
H.-J. Kreowski et al.
Fig. 10.6 The autonomous unit jobj where j D m1 : : : ml
For each job j D m1 : : : ml the autonomous unit jobj is given in Fig. 10.6. It consists of the rules switch.j; mi / for i 2 f2; : : : ; lg, and remove.j; ml /. The rule switch.j; mi / removes the edge from j to the machine mi 1 , inserts a new edge from j to the machine mi and labels the new edge with pt.mi ; j /. This rule can only be applied if the residual processing time at the edge connecting the job with mi 1 is equal to 1. Hence, switch.j; mi / simulates the last processing step of j on machine mi 1 and the switch to the next machine mi . The rule remove.j; ml / removes the job from the environment if the residual time of j on ml is equal to 1. The control condition of each job unit jobj is a regular expression that requires one application of each switch.j; mi /, for i D 2; : : : ; l in the order switch.j; m2 /; : : : ; switch.j; ml / followed by one application of remove.j; ml /.2 Hence, the control condition requires that the job goes through every one of its machines. The goal of jobj is equal to all.
2
If l D 1, no switch-rule is applied.
10
Communities of Autonomous Units
123
It can be shown that the modeling of JSP by Com .JSP/ is correct in the sense that each successful run models a feasible schedule the makespan of which corresponds to the length of the run. The definition of the corresponding schedule is based on the fact that for task t the rule select.t/ is applied exactly once in every successful run. The other way round, there is a successful run for every schedule of JSP. This is expressed in the next observation. Observation 1. 1. Every successful run run D G1 ) ) Gn of Com.JSP / gives rise to the feasible schedule startrun where:
r1
rn
(a) for each task t; startrun .t/ is the index of the unique (parallel) rule containing select.t/, and (b) makespan.startrun / D n. 2. If start is a feasible schedule of JSP with makespan.start) D n, then there is a successful run run D G1 ) ) Gn with startrun D start. r1
rn
Remark. If all autonomous units of Com .JSP/ are busy, then all tasks of the modeled schedules are processed without interruptions. As mentioned before, an autonomous unit is busy if it applies a rule whenever this is possible without violating its control condition.
10.4 Computational Power and Its Limitations In this section, we discuss the computational power of communities of autonomous units and its limitations. First of all, the communities model interacting processes with an emphasis on logistic applications. If the means of modeling are not restricted, the communities can describe processes far beyond computability as further considered in Sect. 10.4.1. If all the components of communities and autonomous units are restricted to computable items, then the semantics of communities becomes algorithmic and communities can be interpreted as computational models. As such they are bounded by the limits of computability. But there are no further limitations as the communities of autonomous units turn out to be computational complete. This is a consequence of the transformation of Turing machines into communities of autonomous units as constructed in Sect. 10.4.2. A further consequence is deduced in Sect. 10.4.3. It turns out that polynomial communities can compute all problems in the famous complexity class NP of decision problems with nondeterministic and polynomial solutions. This is of great interest in the area of logistics because most logistic planning and optimization problems are closely related to the class NP.
124
H.-J. Kreowski et al.
10.4.1 Beyond Computability On the general level of modeling communities of autonomous units, the framework is very generic and generous allowing one to use descriptions of environments, goals, rule applications and control conditions far beyond computability and even beyond consistency. This may sound a bit strange, but it is meaningful on the level of requirement definitions. A modeler who tries to describe the intentions and purposes of a system of processes should not be forced to think in terms of feasibility. It should be possible to formulate the impossible. For example, if the sites in some map have x- and y-coordinates out of some set of numbers, than it is somewhat reasonable to wish that the shortest distance between two sites can be computed. And it is not very difficult to specify this wish if one is familiar with Pythagoras’ law: the positive square out of the sum of the two squares of the differences of the two x- and the two y-coordinates. This root exists in the real numbers, but they are not computable. The root does not always exist in sets of numbers like integers, rational numbers or fixed-point numbers. The problem can be formulated and is not at all unreasonable, but it is not solvable in an obvious algorithmic setting. There are many more examples like this. The problem is the choice of the language to formulate requirements. One needs some data types and some kind of logic to express desired or undesired properties for the intended processes on the data types. If one chooses both quite restrictive like, for instance, finite types and propositional calculus, then many interesting properties cannot be expressed. If one chooses data types and logic with proper expressibility, then it happens easily that one can express uncomputable processes and inconsistent requirements (without any fulfilling process at all). On the other hand, it should be clear to stick to algorithmic concepts whenever one starts to solve a problem that is meant to run on a computer at the end.
10.4.2 Computational Completeness If all components of a community of autonomous units are of an algorithmic nature, then the interacting processes of the community are restricted to the limits of computation. In other words, distribution, cooperation and autonomy do not provide additional computational power. But the good news is that one does not need to take further limitations into account. To show this, Turing machines are transformed into communities of autonomous units. Due to Turing’s Thesis, Turing machines are computationally complete meaning that every binary relation on the set A of all strings over some alphabet A which can be computed in some meaningful way can also be computed by a Turing machine. Since Turing machines can be modeled as communities of autonomous units, the latter framework is computationally complete, too.
10
Communities of Autonomous Units
125
A Turing machine is a system TM D .Z; A; B; d; s0 ; F / consisting of a finite set Z of states, a working alphabet B that contains the alphabet A and some further auxiliary symbols including the box , a state transition relation d Z B Z B fn; l; rg, an initial state s0 2 Z, and a set of final states F Z. To run its computations, it has a finite band of cells each inscribed by some symbol in B where the box represents empty cells. Moreover, there is a read/write-head. At the beginning, an input word w D a1 : : : an 2 A is written on the band cell by cell from left to right. The head is placed on the left-most cell, and the machine is in its initial state. A computation step depends on the current state s and the symbol a in the cell of the head. If .s; a; s 0 ; b; m/ 2 d , then the follower state is s 0 , the head overwrites a by b and moves due to m. If m D l, then it moves to the left neighbor cell. If m D r, then it moves to the right neighbor cell. And if m D n, then it stays on the current cell. If the head should move beyond the left or right end of the band, then new empty cells are added. A computation consists of an arbitrary sequence of such steps. A computation may stop whenever a final state is reached. Then the longest word w0 2 A on the band that starts with the symbol under the head is considered as output and the pair .w; w0 / of input and output words belongs to the computed relation rel.TM/. A Turing machine TM D .Z; A; B; d; s0 ; F / can be modeled as a community of autonomous units Com .TM/. Each environment consists of the disjoint union of the state graph stgr.TM; sN/ with a current state sN and a graph representation band.u; v/ of the Turing band with the inscription uv and the head on the first symbol of v (or on if v D ). The state graph stgr.TM; sN/ has the node set Z. The state transition relation d is a subset of the set of edges with the first projection as source map, the third projection as target map and the projection to the second, fourth and fifth components as labeling. Moreover, each s 2 Z carries an s-labeled loop, each s 00 2 F a fin-labeled loop in addition, and sN an extra curr-labeled loop. The graph band.u; av/ with u D a1 : : : am and v D b1 : : : bn has the form
while band.a1 : : : am ; / has the form
The initial environments are described by stgr.TM; s0 /Cband.; A / specifying the graphs stgr.TM; s0 / C band.; w/ for all w 2 A . The goal is described by stgr.TM; F /Cband.B ; B / specifying the graphs stgr.TM; s 00 /Cband.u; v/ with s 00 2 F and arbitrary u; v 2 B . The community Com .TM / has an autonomous unit state.s/ for each s 2 Z each allowing one state transition from s to some follower state. This is expressed by the constant control condition once which makes sure that only one of the rules
126
H.-J. Kreowski et al.
is applied whenever the unit gets into action. Moreover, there is a unit bandactions providing the reading, writing, and moving on the band. The units state.s/ and bandactions are depicted in Figs. 10.7 and 10.8 respectively.
Fig. 10.7 Autonomous unit state(s)
Fig. 10.8 Autonomous unit bandactions
10
Communities of Autonomous Units
127
Fig. 10.9 Community Com.TM /
If the head-edge is parallel to an a-edge in the graph band .u; v/, then there is exactly one band action rule applicable for any choice of b and m. In every initial environment, the current state is the initial one, and the state transition rule moves the current state to a follower state so that only the unit of the current state can act at any time. If it applies one of its state transition rules trans.a; s 0 ; b; m/ in parallel to the only possible band action with respect to the parameters .a; s 0 ; b; m/, then this step corresponds exactly to a computation step of the Turing machine. Consequently, the iteration of such derivation steps of the community Com.TM / corresponds to computations of TM. To avoid other derivations, the control condition (state jjd bandactions) is employed. It requires that a run of Com.TM / is an arbitrary iteration of one of the state units together with the unit bandactions where each step is synchronized by the state transition d in the following way: If .s; a; s 0 ; b; m/ 2 d , then the unit state.s/ acts by applying the rule trans.a; s 0 ; b; m/, and, in parallel, the unique band action rule that is applicable with the parameters .a; b; m/ is applied. Altogether, the community Com .TM / is defined in Fig. 10.9 as follows. Turing machines are usually considered as centralized computational models. But our remodeling of Turing machines as communities of autonomous units shows that they can also be interpreted as networks of decentralized actors working together on a shared memory in a certain way.
10.4.3 Polynomial Communities The complexity class NP contains all decision problems that can be solved by nondeterministic algorithmic processes in a number of atomic steps which is polynomial in the size of the input. The class NP is quite troublesome because all known solutions of its hardest problems, the so-called NP-complete problems, are exponential, and it is an open problem whether efficient polynomial solutions exist at all. This famous open P D NP-problem is extremely important for logistics because most logistic problems concerning various kinds of route planning and machine scheduling (and many other like these) are NP-hard meaning that they are as difficult as
128
H.-J. Kreowski et al.
or even harder to solve than NP-complete problems. Therefore, most planning and optimization in logistics are based on some heuristics or approximation rather than on exact solutions. The computations of communities of autonomous units are nondeterministic in general because in one step several units may be able to apply a rule and each unit may be able to apply several rules or one rule at several matches. As a rule application is performed in polynomial time (at least in the approach introduced in Sect. 10.2), the computations of a community become polynomial if the lengths are polynomially bounded and all involved conditions can be checked in polynomial time. Consequently, such a polynomial community solves a problem in NP if an initial graph yields the answer YES, whenever there is a computation into an environment that satisfies the goal of the community, and NO otherwise. The situation with respect to the class NP is similar to the one concerning the general computability as discussed in the previous section. Polynomial communities face the same troubles as other approaches to solve NP-problems meaning that autonomy, distribution and interaction do not provide extra help in solving such problems in general. But again there are no further limitations. Or positively stated, all NP-problems can be solved by polynomial communities. To prove this result, one can use the transformation of Turing machines into communities of autonomous units in Sect. 10.4.2. The class NP can formally be defined as the set of all decision problems that can be solved by polynomial Turing machines. A Turing machine is polynomial if the number of steps is bounded by a polynomial in the length of the input word. As each step of a Turing machine is simulated by a rule application in the corresponding community, the latter is also polynomial.
10.5 Conclusion In this paper, we have presented a short introductory survey of communities of autonomous units as graph- and rule-based devices to model interactive logistic processes and discussed some computational aspects of this framework. It has turned out that the framework is computationally complete and covers the famous complexity class NP if polynomial communities are considered and decision problems solved. The bad news is that interaction and autonomy do not provide any extra computational power in these cases. Interaction and autonomy are very meaningful from the modeling point of view in logistics because logistic networks become more and more complex and dynamic as well as several parties may be involved who are not willing to share all their information. But further research is expected to reveal that interaction and autonomy may also be useful to improve the computational conditions. Massive parallelism in the interaction of units may be one promising topic. The use of proper heuristics in the autonomous decisions of unit may be another helpful perspective.
10
Communities of Autonomous Units
129
References 1. Bauer B, Müller JP, Odell J (2001) Agent UML: A formalism for specifying multiagent software systems. Int J Softw Eng Knowl Eng 11(3):207–230 2. Blazewicz J, Ecker K, Pesch E, Schmidt GJW (eds) (2007) Handbook on scheduling: from theory to application. Springer 3. Brucker P (2009) Job-shop scheduling problem. In: Floudas CA, Pardalos PM (eds) Encyclopedia of optimization, 2nd edn., 1782–1788. Springer 4. Corradini A, Ehrig H, Heckel R, Löwe M, Montanari U, Rossi F (1997) Algebraic approaches to graph transformation part I: Basic concepts and double pushout approach. In: Rozenberg G (ed.) Handbook of graph grammars and computing by graph trans-formation, vol. 1: Foundations, 163–245. World Scientific, Singapore 5. Dashkovskiy S, Kreowski H-J, Kuske S, Mironchenko A, Naujok L, von Totth C (2010) Production networks as communities of autonomous units and their stability. Int Electron J Pure Appl Math 2(1):17–42 6. Depke R, Heckel R, Küster JM (2002) Formal agent-oriented modeling with UML and graph transformation. Sci Comput Program 44(2), 229–252 7. Gehrke J, Herzog O, Langer H, Malaka R, Porzel R, Warden T (2010) An agent-based approach to autonomous logistic processes. Künstliche Intelligenz 24(2):137–141 8. Habel A, Heckel R, Taentzer G (1996) Graph grammars with negative application conditions. Fundamenta Informaticae 26(3,4):287–313 9. Hölscher K, Klempien-Hinrichs R, Knirsch P, Kreowski H-J, Kuske S (2007) Autonomous units: Basic concepts and semantic foundation. In: Hülsmann M, Windt K (eds) Understanding autonomous cooperation and control in logistics – the impact on management, information and communication and material flow, 103–120. Springer, Berlin Heidelberg, New York, USA 10. Hölscher K, Knirsch P, Luderer M (2008) Autonomous units for communication-based dynamic scheduling. In: Haasis H-D, Kreowski H-J, Scholz-Reiter B (eds.) Dynamics in logistics, proceedings of the first international conference on dynamics in logistics (LDIC 2007), 331–339. Springer 11. Hölscher K, Kreowski H-J, Kuske S (2009) Autonomous units to model interacting sequential and parallel processes. Fundamenta Informaticae 92(3):233–257 12. Kreowski H-J, Kuske S (2008) Communities of autonomous units for pickup and delivery vehicle routing. In: Schürr A, Nagl M, Zündorf A (eds.) Proceedings 3rd Intl. Workshop on Applications of Graph Transformation with Industrial Relevance (AGTIVE’07). vol. 5088, 281–296 13. Kreowski H-J, Kuske S (2010) Autonomous units and their semantics – the concurrent case. In: Engels G, Lewerentz C, Schäfer W, Schürr A, Westfechtel B (eds.) Graph transformations and model driven enginering - essays dedicated to manfred nagl on the occasion of his 65th birthday, lecture notes in computer science, vol. 5765, 102–120 14. Kuske S, Luderer M (2010) Autonomous units for solving the capacitated vehicle routing problem based on ant colony optimization. Electron Commun EASST 26:23 15. Pinedo M (2008) Scheduling: theory, algorithms, and systems. Springer 16. Rozenberg G (ed) (1997) Handbook of Graph Grammars and Computing by Graph Transformation, Vol. 1: Foundations. World Scientific, Singapore 17. Scholz-Reiter B, Kolditz J, Hildebrandt T (2009) Engineering autonomously controlled logistic systems. Int J Prod Res 47(6):1449–1468 18. Wooldridge M (2009) An introduction to multiagent systems, 2nd edn. Wiley
Chapter 11
Potentials and Limitations of Autonomously Controlled Production Systems Bernd Scholz-Reiter, Michael Görges, and Henning Rekersbrink
11.1 Introduction Manufacturing enterprises are increasingly challenged by a dynamic and volatile business environment. Driving forces in this context are for example an increasing demand of customers for individualized goods, short delivery times and a strict adherence to due dates. Moreover, internal factors like machine breakdowns or rush orders lead to additional dynamics. In order to sustain competitive, manufacturing enterprises have to react promptly to these changes. Conventional centralized production planning and control methods are not able to cope with these dynamics in an appropriate manner [8]. In this context the application of novel decentralized approaches, like autonomous control, seems to be promising. The concept of autonomous control aims at shifting decision making capabilities form the total system to its elements [24]. This approach enables autonomous decision making of intelligent logistic objects. The term intelligent logistic object is broadly defined and covers material objects (e.g., parts in a shop floor) as well as immaterial object (e.g., production orders). On this basis the concept of autonomous control aims at increasing robustness and performance of logistic systems [25]. Previous studies have shown the effectiveness of autonomous controlled production systems in highly dynamic situations compared to conventional methods. Nevertheless, conventional planning methods tend to outperform autonomous control in well-defined situations with less dynamics [15]. This paper addresses the boundaries and the potentials of autonomous control in different static and dynamic situations. The main hypothesis in this context reads as follows: Autonomous control performs best in complex and dynamic situations, while conventional planning methods outperform autonomous control under less dynamic and static conditions.
B. Scholz-Reiter, M. Görges (B), and H. Rekersbrink Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany e-mail:
[email protected] M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_11,
131
132
B. Scholz-Reiter et al.
variable number of stages
variable number of machines
machine
buffer
source
sink
Fig. 11.1 Scheme of Flexible Flow Shop (FFS) problems [10]
Therefore, this contribution evaluates two different autonomous control methods and twelve conventional scheduling heuristics in a Flexible Flow Shop (FFS) environment. The FFS scheduling problem is a well known and common task in operations research. Basically, the FFS structure which is depicted in Fig. 11.1 comprises several production stages with a certain amount of parallel machines per stage [1]. Jobs running through the FFS system have to pass each production stage once. In addition to the basic version of the FFS problem, practice-oriented extensions with unrelated parallel machines and sequence dependant setup times are considered. The analysis covers scenarios with different degrees of structural complexity and dynamics. In this respect the structural complexity refers to the number of production stages and the number of parallel machines per stage. Moreover, different degrees of dynamics are modeled by varying inter-arrival times of the incoming work load: The dynamic situations are characterized by distributed inter-arrival times of the jobs. By contrast, in the static situation all jobs are released at once. The structure of this contribution is as follows: Sect. 11.2 briefly describes the FFS problem class and its extensions. On this basis Sect. 11.3 presents classical scheduling heuristics for this problem class. Subsequently, Sect. 11.4 provides information about the concept of autonomous control in manufacturing and presents particular autonomous control methods, which are used in this evaluation study. The results of the evaluation study are presented and discussed in Sect. 11.5. Finally, Sect. 11.6 summarizes the potentials and the limitations of autonomous control in manufacturing and gives an outlook with further research fields.
11
Potentials and Limitations of Autonomously Controlled Production Systems
133
11.2 Flexible Flow Shop Problems Within the operation research domain, the FFS scheduling problem is a common and well-known task. Generally, this problem class aims at sequencing and assigning a set of jobs to a set of production resources in an optimal manner [1]. However, for most instances of this problem class optimal solutions cannot be found in an applicable computational time. The main reason for this is the complexity of the problem class, which is NP-hard [3]. Hence, heuristic approaches were developed and applied to these problems in the past in order to derive acceptable solutions. This section gives a detailed description of the structure of this problem class. Subsequently, different solution heuristics are introduced. The FFS problem is characterized by a set of production stages, which are sequentially arranged. Each of these production stages comprises at least one machine or different parallel machines. Generally, a FFS scenario has at least one production stage with more than one machine. Figure 11.1 depicts an example of a FFS scenario. Jobs running through this network have to pass each production stage once. Each job can be processed by every machine on a production stage. A practice-oriented extension of this basic scenario is the extension of unrelated parallel machines and setup times [5]. The term unrelated parallel machines indicates, that the processing times and the setup times on a production stage may vary between the machines for a particular job type. With regard to scheduling problems in practice, this formulation is considered to be realistic. Jungwattanakit et al. (2008) present a problem formulation of the FFS problem, which is adapted for this contribution as follows [6]: Parameters and Variables: J S T Mt rj t sm;u;s psst vtm t pm;s
number of jobs number of job types (indices s and u) number of stages number of parallel machines at stage t (index m) release time of job j setup time from job type u to s at machine m at stage t standard processing time of job type s at stage t relative speed of machine m at stage t t processing time of s on m at t, where pm;s D psst =vtm
Logistic target measures Cj completion time of job j Cmax makespan, where Cmax D max.Cj / throughput time of job j ; where Tj D Cj rj Tj TPT mean throughput time; where TPT D mean.Tj / UTL systems average utilization According to this formulation, there are T stages with M t unrelated machines at each stage t. Moreover, there are S different types of jobs, which have all different
134
B. Scholz-Reiter et al.
t t processing times pm;s and setup times sm;u;s on the machines. Ever job has a predefined release time, denoted by rj .
11.3 Scheduling Heuristics The FFS problem in its basic form is NP-hard [9]. Accordingly, optimal solutions can only be determined for very small and less complex scenarios in an applicable computational time. Many different heuristic approaches were developed in the past, to generate suitable solutions for the FFS problem. As far as the FFS with the extension of unrelated parallel machines and setup times is concerned, more sophisticated heuristics are necessary. Jungwattanakit et al. (2005) introduces a heuristic approach containing a set of constructive algorithms and a greedy algorithm [5]. All constructive algorithms are based on heuristic for the flow shop scheduling. They construct an initial sequence for the first production stage. Jungwattanakit et al. (2005) propose the following sequencing algorithms for the first stage: PAL (a slope index heuristic by Palmer), CDS (a best choice heuristic by Campbell, Dudek and Smith), GUP (a slope index heuristic by Gupta), DAN (a heuristic by Dannenbring), NEH (a constructive heuristic by Nawaz, Enscore and Ham) [5]. Subsequently, a greedy algorithm allocates the jobs to the machines at the stages. It assigns jobs to machines considering the setup state of a machine, the processing time of a job on the machine and the idle time of the machines. This algorithm repeats for every stage of the scenario, until a complete schedule is generated for all stages. For a detailed description of the constructive and the greedy algorithm can be found in Jungwattanakit et al. (2005) and Jungwattanakit et al. (2008) [5, 6]. In order to improve the results of this basic procedure with construction of an initial sequence, which is followed by an incremental assignment to machines, Jungwattanakit et al. (2009) propose a coupling of this basic procedure with a genetic algorithm (GAL) [7]. This algorithm takes the sequence for the first stage as genomes. It creates an initial population out of the results of all basic heuristics. The optimization results of the genetic algorithm outperform the solutions of the basic procedure. Thus, the GAL is taken as reference benchmark of all scheduling heuristics in this evaluation study. Due to the expected shortcomings of these scheduling heuristics in dynamic situations, Scholz-Reiter et al. (2010) developed adapted versions of these heuristics for rolling horizons (rPAL, rCDS, rGUP, rDAN, rNEH and rGAL) [18]. The rolling horizon scheduling heuristics (rSCD) divide the entire planning horizon chronologically into sub-planning-horizons, which comprise 25 jobs each. These sub-instances are solved sequentially. The final schedule is generated by merging the sub-schedules. Additionally, a further benchmark is considered. It is the lower bound of the evaluation. This boundary contains the theoretical minimum values, which can be archived in a scenario. For Cmax minimum is determined by:
11
Potentials and Limitations of Autonomously Controlled Production Systems
min Cmax D max
t t max.r / Cmean.sm;u;s // j / C T .mean.pm;s t mean.s / m;u;s t /C J mean.pm;s Mt
135
!
Similar to Cmax the minimum of the TP T is estimated: t t min TPT D T mean.pm;s / C T mean.sm;u;s /
In contrast to the Cmax and the TP T , the best possible utilization of a scenario is 100%. The lower bound is used for estimating the quality of the simulation results. Note, that these general lower bounds are theoretical values. It is assumed, that there is no feasible solution for realizing these lower bounds in the particular scenarios.
11.4 Autonomous Control in Manufacturing The Collaborative Research Centre 637 “Autonomous Cooperating Logistic Processes – a Paradigm Shift and its Limitations” gives the following comprehensive definition of autonomous control: “Autonomous control describes processes of decentralized decision-making in heterarchical structures. It presumes interacting elements in non-deterministic systems, which possess the capability and possibility to render decisions independently. The objective of autonomous control is the achievement of increased robustness and positive emergence of the total system due to distributed and flexible coping with dynamics and complexity.” [24]. According to this definition, the concept of autonomous control is characterized by a shift of decision making capabilities to the logistic objects. Generally, this concept can be applied to various logistic areas, like the intelligent container [4], transport planning [11, 12, 19], manufacturing systems [13] or entire production networks [16]. In the context of manufacturing systems the implementation of autonomous control aims at enabling jobs to find routes through a production system referring to their own logistic targets. This kind of autonomous decision making and the corresponding interactions between the logistic objects aims at a generating a self-organizing behavior which increases the robustness and the performance of the system. This self-organisation is a called emergent behaviour of a complex dynamic system and not derivable from single characteristic [21–23]. Different autonomous control methods were developed in the past. These methods have shown promising results concerning the system’s ability to cope with dynamics and unforeseen disturbances e.g. [13]. Scholz-Reiter et al. (2009) compared different autonomous control methods with a conventional production planning and control (PPC) approach in a real data based model of a production system [15]. It was shown that different autonomous control methods outperform the PPC approach in a highly dynamic situation. Moreover, this study confirmed that different autonomous methods lead to variations in the systems performance, depending on logistic target prioritization.
136
B. Scholz-Reiter et al.
Scholz-Reiter et al. (2010) propose a classification of different autonomous control methods according to their information horizon [18]. It refers to the approach of collecting and processing necessary information for decision making of autonomous control methods. This differentiation identifies local information methods and information discovery methods. All methods of both groups enable local, autonomous decisions of jobs, but they use different information horizons for the decision making process. Local information methods collect information exclusively from the direct neighborhood (buffers or machines). By contrast, information discovery methods collect selected information from the entire production system. Usually, these methods do not discover all available data. It is rather directed to all relevant information.
11.4.1 Local Information Methods Local information methods avoid elaborate discovery procedures. Generally, they aim at rendering suitable local decision with less computational effort. All local information methods have in common that they focus only on information, which are available from the direct local environment (e.g., data form succeeding buffers and machines). The underlying decision procedures are rather simple, compared to information discovery methods. According to a classification, introduced by Windt and Becker (2009), these local information methods can be grouped as follows: rational strategies, bounded rational strategies, combined strategies [26]. This classification is based on the underlying decision mechanisms used by the different autonomous control methods. Rational strategies utilize rational measures for decision making. This means for example estimated buffer and waiting times or due dates of orders. The queue length estimator method (QLE) is a rational strategy. Jobs using the QLE method are able to collect information about the states of all direct succeeding alternative production resources. These jobs estimate the respective processing and waiting times. For further processing, the respective job will choose the alternative with the lowest estimated waiting and processing time [13]. By contrast, biologically inspired methods belong to the bounded rational strategies. They aim at transferring fundamental mechanisms of biologic self-organizing systems to autonomous decision making methods. There are methods, which use the foraging behavior of ants and honey bees or the chemotaxis movement principles of bacteria [2, 14, 17]. In order to keep the evaluation study comprehensible, the QLE method is implemented to the scenarios of the FFS as a representative of the local information methods.
11.4.2 Information Discovery Methods In contrast to local information methods, information discovery methods focus on globally distributed information, which is available in the production system. They
11
Potentials and Limitations of Autonomously Controlled Production Systems
Job
Machine
Job
send one RouteRequest
pass information on utilisation, production times, other job routes ...
send one RouteRequest
receive several RouteReplys
send several RouteAnnouncements
receive several RouteReplys Route Decision
Route Decision [RouteDisAnnouncements]
137
Collect and Update information on job routes
[RouteDisAnnouncements]
send several RouteAnnouncements
Fig. 11.2 Schematic interactions between jobs and machines [12]
conduct higher computational efforts to collect data from different network stages. The Distributed Logistic Routing Protocol (DLRP) belongs to the class of information discovery methods. In its origins, the DLRPt was designed for transport logistics [12]. The index t indicates its design for transport logistic purposes. This protocol is inspired by communication algorithms in wireless ad hoc networks. In the transport logistic context, the DLRPt enables autonomous routing of goods and the corresponding vehicles. However, the application of these principles is not limited to autonomous control of transport logistic processes. Rekersbrink et al. (2010) transferred the approach of the DLRP to autonomous control of shop floors environments [12]. The DLRPp for production systems has two different object types: machines and jobs. Figure 11.2 illustrates the interactions between both object types within the DLRPp . In order to find a route through a production system, a job sends a route request to all succeeding machines. The machines fill all necessary information, like waiting times, setup states or urgencies, into the route request. Subsequently, they pass the request to all further successors. This process repeats until the last production stage is reached. The last machine sends back the information as route reply to the job. On this basis, the job is able to receive the route replies and to select a route according to its individual target preferences. After deciding for one or more routes the job sends route announcements to the machines involved. In contrast to classical scheduling methods, the DLRPp supports a real-time autonomous decision making process, which goes beyond pure pre-planning procedures. Hence, jobs are able to revise previous decisions in terms of generating route disannouncements and new route announcements. Furthermore, the second object class (machines) is able to select jobs form the buffer according to their target preferences.
138
B. Scholz-Reiter et al.
Due to the complexity of the DLRPp , this contribution gives only a brief description of the protocol. For a detailed description of the DLRPp for production systems see [12]. In this contribution the DLRPp is modeled as a representative of the information discovery methods. Jobs using the DLRPp will base their routing decision on the expected completion time. Each job sends two route announcements with different preference values. Additionally, the machines select jobs from the corresponding buffer according to the shortest setup time rule.
11.5 Evaluation and Results Different instances of the FFS scheduling problem are used to compare the performance of the classical scheduling heuristics, the local information method and the information discovery method concerning the structural complexity and different degrees of dynamics. The following presents the concrete parameterization of the instances used. Subsequently, the simulation results are presented and discussed.
11.5.1 Problem Instances The problem instances used are based on the “large size instances” introduced by [5]. There are S D 10 different job types. Every instance contains J D 250 jobs, in total. A uniform distribution is used for assigning jobs to job types. A uniform distribution is used as well, to obtain the standard processing times, the relative speed of the machines and the setup times. Table 11.1 summarizes the system related time parameters and the corresponding interval. The structural configuration of the scenarios are varied. There are instances with T D 3; 5 and 10 production stages. Moreover, the number of parallel machines per stage is varied. Accordingly, there are instances with M t D 3; 5 and 10 machines per stage. In order to model different degrees of dynamics Sung and Kim (2002) used a uniform distribution with varying interval ranges [20]. Similar to this approach, the
Table 11.1 Parameterization of instances Parameter Description t sm;u;s Fixed psts Fixed Fixed vtm T Variable for scenarios Mt Variable for scenarios Variable for scenarios
Distribution Uniform distribution Uniform distribution Uniform distribution Const. Const. Exponential distribution
Interval/values [1 50] [1 10] [0.7 1.3] {3, 5, 10} {3, 5, 10} {st, 0.5, 0.1, 0.075, 0.05}
11
Potentials and Limitations of Autonomously Controlled Production Systems
139
inter-arrival times of jobs rj C1 rj are set to an exponential distribution with a mean value . Consequently, the arrival process is a poisson process. In order to model different degrees of dynamics, is varied form D st; 0:5; 0:1; 0:075 to 0.05. Lambda denotes dynamic aspects of the incoming work load. In the static situation ( D st, this means 1/ all jobs have the release time rj D 0, which means that all jobs enter the system simultaneously. By contrast, a decrease of leads to more extended arrival interval. Thus, the workload of the system depends on .
11.5.2 Evaluation Figure 11.3 presents the results concerning Cmax of both autonomous control methods and for the genetic algorithm (GAL). Each graph of Fig. 11.3 shows the results of Cmax in a particular scenario configuration against the different degrees of
10 stages Cmax [time units]
5 stages Cmax [time units]
3 stages Cmax [time units]
3 parallel machines
5 parallel machines
10 parallel machines
6000
4000
2000
6000
4000
2000
6000
4000
2000
st
0.5
0.1
0.075 0.05 st
l [orders/time unit]
0.5
0.1
0.075 0.05 st
l [orders/time unit]
QLE method
Fig. 11.3 Simulation results for makespan
DLRP
GAL
0.5
0.1
0.075 0.05
l [orders/time unit] min Cmax
140
B. Scholz-Reiter et al.
dynamics. In addition, each graph contains the theoretical minimal values for the respective scenario. According to Fig. 11.3, the performance of the scheduling method (GAL) and both autonomous control methods depends on the degree of dynamics. In all cases, the GAL outperforms the QLE method and the DLRP in the static ( D st) and the nearly static situation ( D 0:5). Here, the GAL performs best. However, with increasing the performance of both autonomous control methods gets closer to the performance of the GAL. Between D 0:5 and D 0:1 there is an intersection of the autonomous control result curves and the GAL curve. When comparing the simulation results with the theoretical minimal values another effect can be observed. The gap between the minimum and the obtained results becomes smaller with an increasing degree of dynamics. For example, in the scenario with five stages and three parallel machines the gap is 21.46% for the GAL (36.82% for the QLE and 48.79% for the DLRP) for D st. By contrast, the gap is for D 0:075 for the GAL 0.55% (0.08% for the QLE and 0.2% for the DLRP). This can be explained by the incoming workload, which differs according to . For lower values of the mean of inter-arrival times get bigger. Hence, the system is under-utilized in this area. Figure 11.3 shows that this effect occurs in every scenario. This effect is more dominant in scenarios, which offer more parallel machines (last column of graph in Fig. 11.3). Due to this utilization effect, the performance of the autonomous control methods and the scheduling methods gets similar. Nevertheless, the explained connection between the methods performance and the dynamics persists, but with smaller differences. These effects are not limited to the results of Cmax . The systems utilization shows similar results. Figure 11.4 depicts exemplarily the utilization of the scenario with ten stages and three parallel machines per stage by showing the composition of the corresponding Cmax values. It differentiates between realized processing times, retooling times and idle times of the machines in the respective simulation runs. The sum of these values is the makespan, which is consequently the height of the bars in Fig. 11.4. Accordingly, Fig. 11.4 give information about the systems utilization, which is the relationship between realized processing times and makespan. Figure 11.4 confirms the observed under-utilized system behavior for small values of . There is a step in the idle times form D 0:075 and D 0:05. This indicates that the system’s capacity is not fully utilized. Furthermore, Fig. 11.4 provides information about the methods ability to minimize processing times, retooling times and idle times. The realized processing times of all methods are very similar for all degrees of , but their performance differs in respect of idle times and retooling times. Especially in the static situation the GAL outperforms both autonomous control methods. It is able to construct schedules with less retooling and idle times compared to the QLE and the DLRP. However, the DLRP leads to the lowest retooling times in the static situation at an expense of longer idle times. The highest degree of retooling times can be found for the QLE method. The same effects can be observed in the nearly static situation D 0:5. Form this point on an increase of leads to rising idle times of the scheduling heuristic. The GAL is not able to assign the incoming workload to the machines in
11
Potentials and Limitations of Autonomously Controlled Production Systems
141
7000 processing time retooling time idle time 6000
5000
4000
3000
2000
1000
0
QLE DLRP GAL
QLE DLRP GAL
QLE DLRP GAL
QLE DLRP GAL
QLE DLRP GAL
λ = st
λ = 0.5
λ = 0.1
λ = 0.075
λ = 0.05
Fig. 11.4 Composition of makespan results of all methods for varying in the scenario with ten stages and three parallel machines
an appropriate manner. It focuses on minimizing the setup times. Thus, the GAL assigns jobs to buffers and machines with jobs of the same type. Consequently, this leads to longer queues, waiting times of jobs and idle times of machines. By contrast, both autonomous control methods perform better in the dynamic situation. Despite longer retooling phases, they are able to harmonize the flow of jobs. Figure 11.5 confirms this. It presents the WIP over time of different simulation runs for the QLE method, the DLRP and the GAL in an exemplary scenario with five stages and ten parallel machines for varying values of . D 0:5; D 0:1 and D 0:075). In the nearly static situation ( D 0:5) all methods collect a high level of WIP up to a certain maximum. At this point, all jobs are assigned to the machines and no new jobs arrive. Accordingly, the WIP level is processed constantly and the WIP decrease with time. In the nearly static situation the incoming work load is bigger than the system’s capacity. This explains results with growing WIP levels. Nevertheless, the GAL performs best in this situation. The GAL decreases the WIP faster than both autonomous control methods. This leads to the lowest Cmax compared to the autonomous control methods. In more dynamic situations ( D 0:1 and D 0:075) the results are different: The GAL still builds up a higher degree of WIP, but the QLE method and the DLRP do not. Both autonomous control methods are able to distribute the incoming work load more evenly to the available machines. Nevertheless, the GAL leads to the shortest Cmax for D 0:1. With regard to the realized WIP levels, both autonomous control methods perform better in this situation. This effect can be observed more
B. Scholz-Reiter et al.
WIP [orders]
142
QLE, λ = 0.5 QLE, λ = 0.1 QLE, λ = 0.075
150 100 50
WIP [orders]
0
0
1000
2000 time [time units]
4000
DLRP, λ = 0.5 DLRP, λ = 0.1 DLRP, λ = 0.075
150 100 50 0
0
1000
2000 time [time units]
3000
4000
GAL, λ = 0.5 GAL, λ = 0.1 GAL, λ = 0.075
150 WIP [orders]
3000
100 50 0
0
1000
2000 time [time units]
3000
4000
Fig. 11.5 WIP over time for the QLE method, the DLRP and the GAL
clearly for D 0:075. Here, the WIP levels and Cmax are lower than those of the GAL. In this particular scenario, the QLE leads to the shortest Cmax for D 0:075. But in general both autonomous control methods perform quite similar. The difference of Cmax between the DLRP and the QLE method is 6.01% for D 0:075. These results show that the classical scheduling heuristics are appropriate in static situation, which tend to be over-utilized. While autonomous control methods are applicable in more dynamic situations. Figure 11.6 strengthens these findings. It presents the average job related throughput times (TPT) for all scenarios and all methods. For reasons of comparability Fig. 11.6 contains additionally information about the lowest possible TPT. Concerning the TPT, the GAL has naturally some shortcomings. Due to the underlying heuristics, the GAL does not focus on optimizing the TPT. By contrast, the introduction of rolling planning horizons helps to overcome these shortcomings. Thus, Fig. 11.6 presents additionally the results of the rolling horizon genetic algorithm (rGAL), in order to give a comprehensible comparison.
11
Potentials and Limitations of Autonomously Controlled Production Systems 5 parallel machines
3 stages mean TPT [time units] 5 stages mean TPT [time units]
3000
10 stages mean TPT [time units]
3 parallel machines 3000
3000
143
10 parallel machines
2000
1000
0
2000
1000
0
2000
1000
0
st
0.5
0.1
0.075
0.05 st
λ [orders/time unit]
QLE method
0.5
0.1
0.075
λ [orders/time unit]
DLRP
GAL
0.05 st
0.5
0.1
0.075
0.05
λ [orders/time unit]
rGAL
min TPT
Fig. 11.6 TPT results of all scenarios
Figure 11.6 confirms the finding stated before: in the static and nearly static situation the GAL performs best concerning the achieved TPT. Similar to the results of Cmax , the gap between both autonomous control methods and the GAL is narrower for scenarios with less production stages (T D 3) compared to scenarios with more production stages (T D 10). There is a huge deviation between the results of the GAL, QLE method, the DLRP and the theoretical minimum values in the static situation. Again, this confirms that the system is over-utilized in this area. This changes with increasing dynamics. In the most scenarios the TPT of the GAL remains at the same level or even increases slightly with . This can be explained by the optimization objectives of the GAL described above. Its primary target is to optimize the makespan. The TPT is a subordinate objective for the GAL. The rGAL performs better than the GAL concerning the TPT in this dynamic situation. Due to the segmentation of the planning horizon into sub horizons, the rGAL optimizes the makespan of
144
B. Scholz-Reiter et al.
the sub-horizons, which leads to a lower TPT in the total schedule. By contrast the rGAL performs even worse in the static situation. Especially in scenarios with 10 stages this effect can be clearly seen. The lowest deviation between obtained results and theoretical minimum in the static situation achieves the GAL. The rGAL performs better with an increase of : A step can be observed in the TPT achieved by the rGAL between D 0:5 and D 0:075. A similar effect is obtained for both autonomous control methods. This indicates, that the utilization state of the system changes in this area from over-utilized to under-utilized situation. Accordingly, there are shorter queues at the machine buffers, which leads to a lower TPT and a more harmonic material flow. Comparing both autonomous control methods and the rGAL in the static situation one can notice that in all cases at least one of the autonomous control methods outperforms the rGAL. Especially, in the scenarios with a higher degree of structural complexity (for example T D 10 and M t D 10) the rGAL performs worse. Due to the segmentation and the partial construction of the schedules, the rGAL performs worse for a lower degree of utilization. In this dynamic situation both autonomous control methods perform best. Furthermore, structural scenario configuration has a curial impact on the performance of the rGAL. With an increasing number of production stages the gap between the theoretical minimum values and the TPT of the rGAL grows. This structural impact is less in the case of autonomous control. In the dynamic situation the TPT of both autonomous control methods remains on a constant level with a smaller gap to the theoretical minimum value.
11.5.3 Conclusions The results concerning Cmax , processing times, retooling times, WIP and TPT show that for the investigated scenarios different scheduling and control methods are suitable. In very static situations with an incoming workload, which is above the system’s capacity, conventional scheduling methods perform best. The GAL provides superior results concerning Cmax and TPT in these situations, independent of the scenarios configuration (number of stages or number of parallel machines). However, with an increase of dynamics both autonomous control methods perform better regarding Cmax and the TPT. Especially, the results of the TPT reveal shortcomings of the GAL to schedule an appropriate flow of jobs through the system. Furthermore, its TPT performance decreases with increasing dynamics in some cases. The extension with rolling horizons (rGAL) attenuates these shortcomings in the TPT performance. Despite, both autonomous control methods outperform the rGAL in dynamic situations. In a more general context, these results confirm the hypothesis stated in the outline: The scheduling algorithms perform best in a static situation, where all information are available. In this static context autonomous control reveals short comings. Due to the idea of autonomous control, which is characterized by interactions and autonomous decision making of intelligent logistic objects in dynamic production environments, this concept is not suitable to
11
Potentials and Limitations of Autonomously Controlled Production Systems
145
static situations. According to the results of this contribution the main application potential of autonomous control is in complex and dynamic systems. A comparison of the autonomous control methods shows a similar performance of both methods. The QLE method performs slightly better than the DLRP concerning the makespan and the TPT, while the DLRP leads to shorter retooling sequences. Due to its ability to discover very complex interconnected production systems it can be expected, that the DLRP performs better in more complex scenarios. This refers to scenarios like sparse production networks. In such networks routing decisions cannot be revised at any stage. Due to the information discovery and the corresponding forecast of succeeding states, the DLRP seems to be favorable in these cases.
11.6 Summary and Outlook This contribution investigated the potentials and the limitations of autonomous cooperating logistic processes in production environments. Therefore, the FFS problem formulation was chosen and different centralized scheduling algorithms for this problem class were introduced. Besides these scheduling algorithms, the evaluation includes two different autonomous control methods. Both methods differ in their information acquirement and information processing procedures. The QLE methods focus solely on local information, while the DLRP is able to discover information distributed in the shop-floor environment. In order to evaluate the scheduling heuristics and the autonomous control methods in respect of their performance in different dynamic situations a set of simulation experiments was defined. Varying degrees of dynamics were modeled by the arrival process of the incoming work load. With regard to variations in the dynamics, the hypothesis stated in the outset was confirmed: the application of autonomous control is limited to dynamic scenarios. In static scenarios classical scheduling heuristics perform best. In situations with an incoming work load, which is higher than the system capacity, classical algorithms are superior to autonomous control methods. By contrast, autonomous control deploys its potential in dynamic situations. Compared to the scheduling algorithms both autonomous control methods allow a continuously flow of jobs through the system, while the scheduling algorithms build up high WIP levels over time. These results encourage for further research in this area. In the next steps of the evaluation of autonomous control extensions like machine breakdowns or rush orders will be addressed. In particular, the effects of this extension on the performance of autonomous control and scheduling methods will be investigated in order to identify further limitations and potentials of autonomous control in production environments. Another research area is the evaluation of autonomous control in combined production and transport scenarios.
146
B. Scholz-Reiter et al.
References 1. Allahverdi A, Ng CT, Cheng TCE, Kovalyov M (2008) A survey of scheduling problems with setup times or costs. Eur J Oper Res 187(3):985–1032 2. Armbruster D, de Beer C, Freitag M, Jagalski T, Ringhofer C (2006) Autonomous control of production networks using a pheromone approach. Phys A 363(1):104–114 3. Garey MR, Johnson DS (1979) Computers and intractability: a guide to the theory of NPcompleteness. Freeman, San Francisco 4. Jedermann R, Moehrke A, Lang W (2010) Supervision of banana transport by the intelligent container. In: Kreyenschmidt J (ed) Coolchain-management. 4th International Workshop. University Bonn, Bonn, 2010, pp 75–84 5. Jungwattanakit J, Reodecha M, Chaovalitwongse P, Werner F (2005) An evaluation of sequencing heuristics for flexible flowshop scheduling problems with unrelated parallel machines and dual criteria. Preprint series: 05–28 (MSC: 90B35) Otto-von-Guericke-Universität Magdeburg Germany 6. Jungwattanakit J, Reodecha M, Chaovalitwongse P, Werner F (2008) Algorithms for flexible flow shop problems with unrelated parallel machines, setup times, and dual criteria. Int J Adv Manuf Technol 37(3):354–370 7. Jungwattanakit J, Reodecha M, Chaovalitwongse P, Werner F (2009) A comparison of scheduling algorithms for flexible flow shop problems with unrelated parallel machines, setup times, and dual criteria. Comput Oper Res 36(2):358–378 8. Kim JH, Duffie NA (2004) Backlog control for a closed loop PPC system. Ann CIRP 53(1):357–360 9. Pinedo ML (2008). Scheduling – theory, algorithms, and systems. Springer, New York 10. Quadt D, Kuhn H (2007) A taxonomy of flexible flow line scheduling procedures. Eur J Oper Res 178(3):686–698 11. Rekersbrink H, Makuschewitz T, Scholz-Reiter B (2009) A distributed routing concept for vehicle routing problems. Logist Res 1(1):45–52 12. Rekersbrink H, Scholz-Reiter B, Zabel C (2010) An autonomous control concept for production logistics. In: Dangelmaier W et al (eds) Advanced manufacturing and sustainable logistics, pp 245–256. Springer, Berlin 13. Scholz-Reiter B, Freitag, M, de Beer C, Jagalski T (2005) Modelling and analysis of autonomous shop floor control. Proceedings of 38th CIRP International Seminar on Manufacturing Systems. Florianopolis, Brazil 14. Scholz-Reiter B, Jagalski T, Bendul J (2007) Autonomous control of a shop floor based on bee’s foraging behaviour. In: Haasis HD, Kreowski HJ, Scholz-Reiter B. (eds) First International Conference on Dynamics in logistics. LDIC 2007. Springer, Berlin, Heidelberg, pp 415–423 15. Scholz-Reiter B, Görges M, Philipp T (2009) Autonomously controlled production systems – influence of autonomous control level on logistic performance. CIRP Ann Manuf Technol 58(1):395–398 16. Scholz-Reiter B, Dashkovskiy S, Görges M, Naujok L (2010) Stability analysis of autonomously controlled production networks. Int J Prod Res. http://www.informaworld.com/ 10.1080/00207543.2010.505215 (accessed 07 December 2010) 17. Scholz-Reiter B, Görges M, Jagalski T, Naujok L (2010) Modelling and analysis of an autonomous control method based on bacterial chemotaxis. 43rd CIRP International Conference on Manufacturing Systems 2010 (ICMS 2010). Neuer Wissenschaftlicher, Wien, pp 699–706 18. Scholz-Reiter B, Rekersbrink H, Görges M (2010) Dynamic flexible flow shop problems – scheduling heuristics vs. autonomous control. CIRP Ann Manuf Technol 59(1):465–468 19. Schönberger J, Kopfer H (2009) Online decision making and automatic decision model adaptation. Comput Oper Res 36(6):1740–1750 20. Sung CS, Kim YH (2002) Minimizing makespan in a two-machine flowshop with dynamic arrivals allowed. Comput Oper Res 29(3):275–294
11
Potentials and Limitations of Autonomously Controlled Production Systems
147
21. Ueda K, Hatono I, Fujii N, Vaario J (2000) “Reinforcement learning approaches to biological manufacturing systems”. CIRP Ann Manuf Technol 49(1):343–346 22. Ueda K, Markus A, Monostori L, Kals HHJ, Arai T (2001) Emergent synthesis methodologies for manufacturing. CIRP Ann Manuf Technol 50(2):535–551 23. Vaario J, Ueda, K. (1997) An emergent modelling method for dynamic scheduling. J Intell Manuf 9(2):129–140 24. Windt K, Hülsmann M (2007) Changing paradigms in logistics – understanding the shift from conventional control to autonomous cooperation and control. In: Hülsmann M, Windt K (eds) Understanding autonomous cooperation & control – the impact of autonomy on management, information, communication, and material flow. Springer, Berlin, pp 4–16 25. Windt K, Böse F, Philipp T (2008) Autonomy in production logistics – identification, characterisation and application. Int J Robot CIM 24(4):572–578 26. Windt K, Becker T (2009) Applying autonomous control methods in different logistic processes – a comparison by using an autonomous control application matrix. Proceedings of the 17th Mediterranean Conference on Control and Automation. Thessaloniki, Greece
Chapter 12
Scalability Effects in Modeling Autonomously Controlled Logistic Processes: Challenges and Solutions in Business Process Modeling Bernd Scholz-Reiter, Daniel Rippel, and Steffen Sowade
12.1 Introduction Today’s logistic systems become more and more complex. In this situation, high fluctuations in customer demand and unforeseen events decrease the predictability of their behavior and increase their dynamics and vulnerability. While classical production planning and control systems are reaching their limits to deal with these effects, autonomously controlled logistic processes are a possible solution [6]. This concept aims to increase a logistic system’s robustness and flexibility by distributing planning and control competencies to logistic objects, e.g. to commodities, halffinished goods, resources, and orders. Autonomously controlled logistic processes rely on the logistic objects’ local decisions and lead to a positive emergence of the overall system’s behavior [6]. Logistic process experts face the tasks to design, model, and evaluate autonomous processes in order to apply autonomous control in logistic systems. This development process includes the specification of logistic processes, logistic objects’ abilities, decision-making strategies, as well as a definition of an overall system in form of a logistic scenario. In order to guide logistic process experts through the development process, a modeling methodology called Autonomous Logistic Engineering Methodology (ALEM) is being developed [20]. The methodology’s applicability and its advantages have been demonstrated at the example of production logistic scenarios like shop-floor manufacturing systems [7]. Beyond that, the advantages of autonomous control obviously increase with the growing size of a logistic system, due to an increasing number of decision alternatives in its running processes as well as during the system’s design process. Consequently, a decentralization of planning and control mechanisms in larger scaled logistic systems results in an increase of flexibility and robustness. For example, supply chains, production networks, or virtual enterprises constitute more complex scenarios than B. Scholz-Reiter, D. Rippel, and S. Sowade (B) Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany e-mail:
[email protected] M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_12,
149
150
B. Scholz-Reiter et al.
simple manufacturing scenarios. They cover a variety of logistic objects, each having a wide range of decision alternatives. An increase in a logistic system’s size and complexity influences the amount of information to be modeled in ALEM. Hence, an interesting question is, if and how this increase affects the modeling process and the ALEM model consequently. ALEM might have to deal with two challenges caused by scalability that emerges during the design and modeling process. First, ALEM’s bottom-up modeling approach requires a detailed modeling of the logistic processes and their components. In large scaled scenarios, this might come along with a low comprehensibility of the models and might face modelers with additional challenges in modeling, testing, enhancing and error tracing. Second, modelers face the task to obtain all information that required for a detailed model of all logistic objects’ abilities, knowledge, processes etc. However, such information might be considered private and might be unavailable for a specific modeler. This is in particular true during the modeling of large scaled logistic systems, which involve a variety of organizational independent entities. Their internal information might not be revealed to the modeler. Hence the question arises, if and how these issues affect the modeling process and its results. Therefore, this article investigates both challenges as they emerge from the application of ALEM to large scaled logistic scenarios. Thereby, it illustrates techniques and methods to cope with them. First, the article presents the paradigm of autonomous control and the corresponding modeling methodology ALEM. Afterwards, it discusses organizational, technical, and personnel issues by making use of two logistic scenarios. The first scenario represents a manufacturing system, while the second illustrates a supply chain. At last, the article sketches approaches to deal with these challenges and limitations.
12.2 The Autonomous Logistic Engineering Methodology: ALEM Hülsmann and Windt define autonomous control as “processes of decentralized decision-making in heterarchical structures. It presumes interacting elements in nondeterministic systems, which possess the capability and possibility to render decisions independently. The objective of Autonomous Control is the achievement of increased robustness and positive emergence of the total system due to distributed and flexible coping with dynamics and complexity” [6]. In order to enable autonomous control, selected logistic objects are equipped with abilities to manage information, to process information (decision-making), and to execute decisions. The heterarchical interactions of these objects form autonomous processes. Although it is impossible to predict the overall system’s performance, simulation studies demonstrate positive effects of autonomous control on the system’s performance, e.g. in terms of logistic goal achievement, flexibility, and robustness [4, 12, 15, 16].
12
Scalability Effects in Modeling Autonomously Controlled Logistic Processes
151
In order to enable logistic process experts to develop and evaluate autonomously controlled logistic systems, the methodology ALEM is being developed. It is designed to provide logistic process experts with tools and methods for the development of autonomously controlled logistic processes. In this role, ALEM satisfies four essential requirements [17]: General requirements: A modeling methodology has to consist of a notation, a procedure model, and of a fundamental structuring that provides orientation during the modeling process. User orientation: A modeling methodology has to be suitable for its users. ALEM focuses on logistic process experts. Therefore, it has to exploit well-known and standardized methods for process modeling in logistics. Domain orientation: A modeling methodology has to be specific to the domain of application. In case of ALEM, this domain is threefold and consists of logistic systems and logistic process modeling, as well as of the paradigm of autonomous control. Model-usage orientation: Created models and parts of them have to be useful in subsequent steps of a development process. Hence, they have to provide essential information about the system in scope, in order to enable its deployment in real world applications. The ALEM methodology provides a view concept and a procedure model, which are described in the next sections, in order to comply with the general requirements. The ALEM’s notation closely conforms to a subset of the Unified Modeling Language (UML). It is a well-known and standardized language that is widely used in the areas of software development (e.g. [2, 9]), knowledge representation (e.g. [26]) and process modeling (e.g. [10]). Hence, it is likely that a potential user of ALEM already had contact with this notation [20]. ALEM extends the UMLNotation by some domain specific diagrams to address the needs of a modeling methodology for logistic systems [8, 20]. For example, these additional diagrams cover the structure of products, which are manufactured within a modeled scenario [22]. Overall, the notation satisfies the user orientation requirement. ALEM follows a bottom-up modeling approach, in order to satisfy the domain requirements on autonomous control. This modeling approach focuses on the logistic objects’ abilities and their processes for decision-making. This focus directly results from the definition of autonomous control. The ALEM methodology spans four major steps of a system’s development cycle (Fig. 12.1) in order to satisfy the model-usage orientation requirement: First, the methodology supports logistic process experts in specifying the logistic processes and the scenario. Second, it supports simulation to evaluate the models. Third, the methodology provides tools and concepts to configure the infrastructure that is necessary to enable the autonomous processes. Finally, the methodology provides means to perform a cost-benefit analysis on its models. Furthermore, the high level of detail, encompassed by ALEM, enables an easy development of an agent-based system out of ALEM models [18].
152
B. Scholz-Reiter et al. Initiation
Pilot-Study
Main Study
Specification and Analysis
Simulation
Cost-BenefitAnalysis
InfrastructureConfiguration
Cancel Detailed Study
Implementation Termination
Introduction
Fig. 12.1 System development cycle [20]
Structure • UML-Class Diagrams • Layout Diagrams
Knowledge • UML Class Diagrams • Product Structure Diagrams • Knowledge Maps
Process • UML–Activity Diagrams • UML–State Machines
Ability • UML-Class Diagrams • Knowledge Maps
Communication • UML– Class Diagrams • UML– Sequence Diagrams
Fig. 12.2 ALEM – view concept [19]
12.2.1 ALEM: View Concept The ALEM view concept [19] distinguishes between five semantic views (Fig. 12.2). These views either describe static or dynamic features of the model. Static features define which intelligent logistic objects exist within the modeled scenario, their knowledge, and their abilities. Dynamic features cover the logistic objects’ processes and communication protocols. Each view may contain aspects referring to either the micro or the macro perspective. Elements within the micro perspective, like decision-making strategies, are object-internal. In contrast, the macro perspective describes object-external elements, such interaction protocols. The semantic views distinguish between structural features (structure view), knowledge aspects (knowledge view), actions which can be performed by the logistic objects (ability view), processes (process view) as well as between communication protocols and message contents (communication view).
12
Scalability Effects in Modeling Autonomously Controlled Logistic Processes
153
The structure view contains structural features of the modeled system. It defines all present logistic objects, as well as relationships between them. In addition, this view includes the spatial layout of modeled scenarios. The knowledge view covers all aspects concerning knowledge and objectives. UML-Class diagrams represent the logistic object’s knowledge in form of attributes. The logistic objects’ objectives are defined in UML-Class diagrams. Additionally, this view incorporates product structure diagrams [22] and knowledge maps. Those maps assign knowledge, modeled within the distinct diagrams, to logistic objects. The ability view uses a UML-Class diagram in order to represent fundamental abilities that logistic objects can perform. Knowledge maps assign these abilities to logistic objects. The process view makes use of UML-State Machines and UML-Activity diagrams in order to represent the behavior of logistic objects. It incorporates object internal as well as system-wide processes. State machines describes logistic objects’ life cycles, while activity diagrams cover complex chain of single activities. The communication view contains UML-Class and UML-Sequence diagrams. UML-Class diagrams define the content of messages, which are exchanged by logistic objects. Sequence diagrams depict communication protocols. The different views’ diagrams are interconnected, through knowledge maps or by direct assignment. In addition to the links indicated as circles in Fig. 12.3, state-machines and activity diagrams refer to other activities, state-machines,
Fig. 12.3 Interconnectivity of ALEM-diagrams (following [20])
154 Fig. 12.4 ALEM – procedure model [7]
B. Scholz-Reiter et al. 1. Objectives 2. Structure 3. Abilities 4. Processes 5. Decision 6. Knowledge
Legend
7. Communication
Feedback Flow General Flow
8. Scenario
Modeling Item
communication protocols, or abilities. This modular modeling approach allows reuse of already present model elements. For example, a set of abilities can be assigned to several logistic objects. The possibility to remove assignments, without removing the modeled element poses another advantage of this modularity.
12.2.2 ALEM: Procedure Model The ALEM Procedure Model [7] closely conforms to the view concepts semantic views (Fig. 12.2). It covers eight steps and follows a bottom-up modeling approach. Feedback loops are allowed between the different modeling steps. Figure 12.4 presents the procedure model and depicts possible interdependencies between its steps. For example, logistic process experts have to ensure that required knowledge is available to logistic objects, if he designs a communication protocol. The first step incorporates a definition of the logistic system’s objectives for each logistic object. For example, these could be high utilization, low throughput times, or low costs. The second step is the definition of a system’s structure. In this step, logistic process experts define which logistic objects will be present in a system as well as their relationships with other logistic objects. As a third step, the experts’ model all abilities, which contribute to a logistic process, and assign them to the logistic objects, defined in the second step. Examples for such abilities are a machines ability to manage its production schedule autonomously or to perform manufacturing steps on commodities. The fourth step includes modeling of complex activities, carried out by the logistic objects. Fifth, logistic process experts design the decision strategy of each logistic object. In the sixth step, they model and assign required knowledge to the logistic objects. In order to enable the logistic objects to render the modeled decisions, the experts design communication protocols that ensure the availability of information at decision-making objects. For example, a
12
Scalability Effects in Modeling Autonomously Controlled Logistic Processes
155
commodity, that decides which machine to select for processing, requires knowledge about processing times, as well as about the next free time slot on a machines schedule. Finally, the last step makes use of the modeled logistic objects, in order to instantiate a specific scenario of the described system in which autonomous logistic processes take place. ALEM has been applied exemplarily to production logistic systems (e.g. [7]). In order to examine implications of larger scaled logistic processes and to the ALEM methodology, the next section briefly introduces two example scenarios of different scale and describes their properties with respect to modeling.
12.3 Scenarios This section introduces a flexible flow-shop manufacturing system and a supply chain scenario. Both scenarios sketch the modeling processes as well as difficulties logistic process experts face during modeling. Building upon these problems, the following section identifies scalability-related limitations on the ALEM methodology.
12.3.1 Flexible Flow-Jobs System Today, flow-shop systems are widely used and usually employ different production and/or assembly stages. Commodities pass each stage in a sequential order given by their product and manufacturing structure. According to Allahverdi et al. [1], flexible flow-job systems use parallel machines on each stage. For reasons of illustration, this article’s example flow-job scenario utilizes the minimum of two production stages, each with two equal machines in parallel (Fig. 12.5). For instance, two turneries form the first stage and two sawmills represent the second stage. In order to apply autonomous control to this flexible flow-job scenario, logistic process experts have to describe the desired autonomous processes. Commodities (e.g. raw materials) enter the production system according to production orders.
Production stage 1
Production stage 2
M1,1
M2,1
M1,2
M2,2
Source
Fig. 12.5 Flexible flow-job scenario [21]
Legend
Mx,y
Sink
Machine y on stage x
Production entry or exit
156
B. Scholz-Reiter et al.
These production orders either originate from external sources, like customers, or arise from internal activities, e.g. for warehousing. With respect to the ordered products’ structure and the objectives associated with the type of order, the commodities proceed through the production stages. On each stage, the commodities select the most suitable machine, according to their objectives. Therefore, they request information from the parallel machines on the next production stage and make their decision. This process reoccurs for each production stage, until the final product finishes manufacturing. In order to model these autonomous processes, logistic process experts first identify those logistic objects that will make and carry out local decisions. In this case, these objects are at least commodities and resources, i.e. machines. Orders may be included as intelligent objects if desired [23]. Following the ALEM bottom up modeling approach, the experts model the objects’ abilities, processes, and knowledge, as well the data exchanges between them. Building upon this basic model of the system, the experts design decision-making strategies, -functions, communication protocols, and include additional knowledge required for decision-making. After modeling the desired system, logistic process experts evaluate the model and refine it if needed.
12.3.2 Supply Chain Scenario While value networks or chains cover all activities that increase a product’s value, i.e. manufacturing and transportation, as well as management activates within and around those organizations related to the product [11], supply chains span the part of a value chain that focuses on the delivery process of products and commodities, as well as on the partners involved. The supply chain example employed in this article originates from a case study conducted in the apparel industry [24]. The case study covers production of garments, their transportation to distribution centers, as well as their shipping to customers. The considered supply chain spans one manufacturing plant, one reloading center, three distribution centers, and several customers. Ships or airplanes perform transportation between manufacturing plant, reloading center, and distribution centers. Motor-trucks serve the customers on the last hop. In each case, different subcontractors perform the transportation task (Fig. 12.6). In contrast to the manufacturing scenario, the desired autonomous processes are more complex within this scenario. They cover manufacturing, different transportation activities, as well as distribution and order assignment. Concerning manufacturing, the decision-making objects are as well commodities, (half) finished goods, and machines. The autonomous processes are quite similar to those described within the last scenario. In addition to these processes, the garments are able to arrange themselves to badges, e.g. packages, palettes, and containers, in order to prepare transportation or distribution. Therefore, the garments coordinate each other and identify conforming delivery dates as well as destinations across orders. The case study dealt with
12
Scalability Effects in Modeling Autonomously Controlled Logistic Processes
Plant
Reloading Center
Distribution Center
Distribution Center
Customer … Customer
157
Distribution Center
Customer … Customer
Customer … Customer Legend
Transportation (Ship) Transportation (Plane)
Processing Center
Fig. 12.6 Supply chain scenario (extending [25])
garments and their autonomous processes, which satisfy customer orders by reference to the number of ordered variants as well as to the garments’ spatial location, and the orders’ priorities. In order to achieve such dynamic behavior, the (badges of) garments decide on their own to decompose into smaller badges or even to single garments in distribution centers, as well as they recombine themselves as demanded in a customer order. The third activity covered by this supply chain scenario is transportation. According to the badges’ or their respective orders’ objectives, transportation devices, like planes, ships, trucks, and trains, are selected autonomously. These transportation devices constitute additional decision-making objects, as they perform routing decisions on behalf of their own or on behalf of (badges of) garments. In order to apply autonomous control to the described supply chain, logistic process experts have to model all decision-making logistic objects, including their knowledge, abilities, decision-making strategies, processes, and data exchanges. Thereby, the organizational independence between manufacturing companies, transportation providers, and distribution center operators leads to difficulties in modeling of all processes and objects in an integrated view. Logistic process experts, belonging to the manufacturing company, are unlikely to have insight into a transportation provider’s processes, inventory, and decision-making strategies and vice versa. This organizational independence further complicates the development of logistic objects that are involved in all of the supply chain’s activities.
12.4 Complexity Induced Challenges An increasing size of the scenario confronts logistic process experts with various challenges. These challenges emerge from the handling of the models and of the methodology. The ALEM methodology contrasts classical process modeling methodologies in its structure. It focuses on the single objects and their partial
158
B. Scholz-Reiter et al.
contributions to the overall process. Therefore, ALEM does not provide a general overview over the overall process, but only provides modeling capabilities for those partial processes of the logistic objects involved. Moreover, the increasing complexity complicates the tasks of creating, analyzing, and modifying ALEM models for logistic process experts.
12.4.1 Local Process Modeling One major challenge is ALEMs focus on single objects’ processes. In contrast to classical process modeling, there exists no general overview, depicting all the process’ activities, regardless of which logistic object executes the activities. A set of diagrams, each referring to single objects’ activities and decision processes, represents a manufacturing process. In order to acquire an impression of the overall process, logistic process experts have to search manually through different diagrams identifying main activities and protocols. The other way round, the bottom-up modeling approach requires logistic process experts to model the intelligent logistic objects’ activities and decision-making strategies first, before combining them to processes that are more complex. This bottom-up way of modeling requires a different understanding of those processes interactions, which is contrastive to the practice of classical process modeling. One option to cope with this challenge could be the introduction of a process design perspective into ALEM’s macro-process view. The respective diagrams would consist of the overall processes main activities, each assigned to a particular type of intelligent logistic object. The assignments indicate which intelligent logistic object performs a specific activity. Transitions could describe information passed between activities. If different intelligent logistic objects perform linked activities, communication protocols are required which convey information or decisions. Figure 12.7 presents a possible example for such an overview. The lack of a general overview as well as a high number of single processes, leads to additional challenges regarding the creation, analysis, and modification of ALEM
Determine Manufacturing Step
Determine possible resources
Determine Manufacturing Resource
Request time slots
Fig. 12.7 Process overview with refinement
Apply Manufacuring Step
Decide possible time slots and reply
Decide for Resource and reply all
12
Scalability Effects in Modeling Autonomously Controlled Logistic Processes
159
models. With an increasing size of a logistic system, the number of decision-making objects and processes increases likewise. On the one hand, this prolongs modeling and testing of the ALEM models. On the other hand, the variety of decision-making functions and activities requires careful modeling to avoid errors.
12.4.2 Model Creation The required effort in creating an initial model increases with a growing number of autonomous logistic processes and objects. This task is complicated by equally typed objects, which apply different sets of decision-making strategies and abilities. For example, within the supply chain scenario, different providers perform transportation on the garments. These transportation processes involve the same types of objects like garments or packages but can apply different control strategies. Therefore, abilities may be distributed differently between intelligent logistic objects of those processes. In such cases, ALEM requires modeling of those objects as individual intelligent logistic objects, although they are of the same type. As a consequence, the complexity of the corresponding ALEM models increases dramatically. Besides the increasing number of intelligent logistic objects to define, the effort in creating a scenario grows. Due to the enlarged system, there exist more intelligent logistic objects, like commodities, resources or devices, that have to be created and set-up during scenario creation. For logistic process experts, it can be difficult to retain an overview over the objects abilities, knowledge, and processes, as well as over the complete scenario and all instances of logistic objects.
12.4.3 Model Modification Applying modifications to existing models poses additional challenges to logistic process experts, when the model’s size and complexity increase. Modifications become necessary when resolving errors, or if experts redistribute knowledge or abilities. In the context of the supply chain example, logistic objects take part in several activities. Each object stays active for a longer period of time and becomes involved in a growing number of successive activities. Within the shop-floor example, the autonomous objects only have to pass manufacturing. Within a supply chain, the objects additionally pass several transportation, packaging, and repackaging stages. Due to the growing number of processes, modifications to a single object’s abilities, communication, or decision-making strategies may affect a variety of processes and thus other logistic objects. Therefore, the interconnectivity between logistic objects grows and an estimation of the impact of local modifications to a single intelligent logistic object, regarding the overall processes, becomes more difficult. For example, modifying a communication protocol in order to reduce the amount of transmitted information can speed up some processes, while leading to a
160
B. Scholz-Reiter et al.
lack of information in others. The redistribution of abilities or knowledge between logistic objects leads to a particular challenge. Redistribution might be necessary to resolve errors or to optimize a system, e.g. if the model expected information that could not be obtained in the real world scenario, of if logistic process experts compare different autonomous control strategies or architectures in order to estimate the suitability of specific configurations for a particular system. Modifications to an object’s abilities or knowledge have to incorporate the overall process’ structure. In case of a logistic process expert optimizing a manufacturing process, he has to ensure, that modified logistic objects still possesses all required abilities and knowledge to perform subsequent processes. Modifications to the object’s knowledge or abilities impact the object’s processes and communication protocols. If knowledge is redistributed to another intelligent logistic object, communication protocols might have to be redesigned. The same accounts for objects’ abilities. In case of modifying an intelligent logistic object’s abilities, it might become necessary to design new processes and protocols in order to maintain a model’s functionality. The ALEM-Procedure Model supports modifying existing models. The procedure model’s feedback loops could be interpreted as validation links, if interpreted in a bidirectional way (compare Fig. 12.4). As an example, modifications to a logistic object’s knowledge require validation of communication protocols; modification to an object’s abilities requires validation of its processes, decisions, knowledge, and communication. In addition, the modeled scenario has to be validated in all cases. Another way to support logistic process experts in modifying an existing model would be the application of automated validation tests. At least on a syntactical level, such tests could identify other segments of the model, which would be rendered inconsistent with the modifications. Amongst others, communication protocols could be tested on modifications to an intelligent logistic object’s knowledge, in order to validate if all knowledge, which should be transmitted, is still present at the object. On changes to an object’s abilities, its processes could be validated to identify those processes that formerly included the modified ability.
12.4.4 Model Testing The great variety of objects and activities prolongs the time until a model can be tested. Further, the duration of simulation runs increases with a growing scenario size. Testing becomes complicated due to the increasing amount of interactions between involved autonomous objects. From a logistic process expert’s point of view, the overall system’s behavior becomes more and more non-deterministic, due to the growing number of interactions and processes. Additionally, the nondeterminism of some autonomous control strategies (e.g. [14]) reinforces this effect. In the context of testing and analyzing the model, non-determinism leads to difficulties in recreating particular situations of interest for the purpose of simulation. In order to test specific situations, logistic process experts have to model them directly
12
Scalability Effects in Modeling Autonomously Controlled Logistic Processes
161
as scenarios. Therefore, logistic process experts have to create a scenario that exactly matches the conditions of the desired situation, involving all objects internal states and positions. With an increasing size of the scenario, more and more objects may be involved in such situations. A related challenge emerges, if logistic process experts observe undesirable behavior. In such cases, they have to isolate those logistic objects, and activities from which the behavior originates. With a growing number of objects, activities and interconnectivity, this task requires a deeper insight into a system and its objects’ information handling and decision-making process. For example, a logistic object’s wrong decision could result from a miss-specified decision function or from the use of outdated or wrong information. A decision function may be a good choice in one situation, while failing in other situations. In case of an object using outdated information, logistic process experts have to evaluate, if the object itself did not properly update the information (internal processes), if it did not request updated information (communication), or in the worst case, if other objects provided outdated information (external source of error). In this case, logistic process experts have to analyze additional logistic objects in order to identify an error’s cause. Once logistic process experts identified the origin of an error, he has to modify and reevaluate the model. Therefore, he has to recreate those situations, in which the object decided wrong. One possibility to cope with this challenge would be the integration of a detailed, simulation-based logging functionality. During simulation runs, a logging component records all decision related information the objects make use of. In addition, it records all information sent and received by the objects. Using this information, logistic process experts can track all information used in a decisions to their origins (Fig. 12.8). If several logistic objects transmit one particular information, a logging component could identify and visualize the information’s route from the decision
Decision: „Decide for Resource“ Timestamp: 19:32
Timestamp: 19:22 Information pool: First empty slot: 34:23
Used Information: : Timestamp 34:23 : Timestamp 24:56
Protocoll: „Request Resource“ Timestamp: 19:22 Content: : Timestamp 34:23
Legend
Decision
Transmission
Fig. 12.8 Example for information tracing
Information at Resource
162
B. Scholz-Reiter et al.
in question to its source. By selecting a decision, the logging component depicts all information used to make the decision as well as the information’s origin and transmission history. Additionally, the logging component could facilitate the creation of specific scenarios, as the state of each object becomes observable at any point in time.
12.5 Organizational Independence Logistics employs several processes in production, assembly, and transportation in order to produce and deliver goods for industrial and end consumers. Usually, different companies perform each of the processes with respect to their specialization for specific products and services which bases on their technological and economical capabilities and resources. Hence, they are organized in form of supply chains, production networks, or virtual enterprises. In this sense, the involved companies are organizationally independent of other companies participating in such networks [3]. For instance, Fig. 12.6 illustrates an apparel supply chain with independently operating processing centers and transportation links. In this case study, one company owns the plant and the distribution centers displayed, while various service providers operate the reloading center and all transportation links [25]. Due to their independence, companies within one supply-chain are unlikely to unveil their internal processes and objectives to their partners in detail. Instead, they specify their demand in form of orders that describe requested goods and services in qualitative and quantitative properties, like dimensions, material, quantity, due date, and location. Thus, independent companies basically interact via contracts and a supply chains’ activities consist of sub-processes, which belong to different partners. However, the manufacturer in the case study is likely unable to influence succeeding transportation processes after handing cargo over to transportation provider. He regains control on delivery at the target location. As shown in the case study, participating companies in such networks are assumed to operate economically independently [3]. They hide their internal operational as well as organizational structure and keep their important business secrets, like politics and strategies. The companies follow their own objectives without accountability for a partner’s behavior and do not have power to direct other partners’ internal processes or structures. Furthermore, their independence proceeds in the methods, standards, and tools used by them internally. Each of them models their part of a supply chain for their own. Nevertheless, functional dependencies can correlate with technological constraints and vice versa, e.g. if a process requires a specific predecessor that is offered only by few other companies. Further, supply chains, production networks, and virtual enterprises form large logistic networks and induce a high level of complexity into the modeling process. Generally, complexity is understood as the quantity of systems elements and their relations to each other. The model complexity of the mentioned networks is higher
12
Scalability Effects in Modeling Autonomously Controlled Logistic Processes
163
than in case of a single manufacturing plant when assuming that such networks consist of several interlinked plants. Autonomous logistic processes lead to additional complexity at the systems elements design and in the dynamic behavior of the overall system [27]. In addition, a collaborative complexity results from the specifics of interaction processes that are required for inter-company cooperation. The system elements, logistic objects, have to be equipped with components that enable decision-making. Dynamic characteristics of the system emerge from fluctuations of quantitative and qualitative parameters in space and time domain. For example, the fluctuations can be distinguished by their linearity, directness. The ALEM modeling methodology uses a view concept in order to cope with an overall models’ complexity. This model complexity results from the amount of elements that need to be modeled, from cooperation-induced issues, from the lack of a global overview about a logistic business process, as well as from difficulties to determine the emergent behavior of systems being under construction. Although the ALEM view concept provides seven distinct dimensions on a model, it excludes illustrations of the overall model to designers, process managers, or single logistic objects, as well as an overview about all information of a selected logistic object. Further, a collaborative view is currently not part of the ALEM methodology. This view can be useful, in order to oversee processes spanning large value networks, supply chains, or virtual enterprises.
12.5.1 Summarizing the Problem Areas An integration of distinct, independent organizations into one logistic system induces several challenges towards modeling of the overall autonomous business processes, if modelers independently use the ALEM methodology. Hence, ALEM has to be enhanced in order to guide multiple modelers at a time through the design and evaluation processes in supply chains, production networks, or virtual enterprises. The first problem area results from a lack of overview about foreign systems in terms of system elements, processes, objectives, etc. Secrecy of this private information causes barriers for the information exchange, which thereon affects the whole modeling process. Logistic process experts can only obtain a comprehensive view on the overall process on an abstract level. However, the ALEM methodology requires modeling of all sub-processes contributing to the overall business process in detail. Logistic process experts face the challenge to model a supply chain partners’ sub-processes as detailed as possible, while specific information is unavailable. For the supply chain example provided, logistic process experts of the manufacturing organization can easily model manufacturing and reloading processes. Processes taking place within a distribution center can be obtained in full detail, as the centers and the manufacturing plant belong to the same organization. In contrast, the manufacturing organization cannot obtain and model transportation processes, as subcontractors carry out these activities independently.
164
B. Scholz-Reiter et al.
The second problem area bases on limitations introduced by the ALEM modeling process. At the moment, the methodology does not include collaboration mechanisms for creation and design of highly autonomous decentralized logistic business processes. Instead, the modeling process is described with the inherent assumption of modeling one logistic business process at a time at one location by a single modeler instance. Logistic process experts are able to model autonomous logistic processes, as long as they take place at the company employing the modeler. This modeling approach works fine if modelers have access to all necessary information. However, it leads to a lower detailed model, if segments of large logistic systems belong to different, organizationally independent companies, e.g. supply chains, production networks, or virtual enterprises whose members model their internal process structures themselves. Hence, ALEM lacks of collaboration mechanisms for modeling and evaluation of autonomous logistic business processes in a decentralized way by different modelers.
12.5.2 Improvements for Collaborative Modeling with ALEM As a consequence of the problem caused by the organizational independence of companies, ALEM has to be modified and must support collaborative modeling of autonomous logistic business processes. For this purpose, the methodology has to adapt suitable mechanisms for collaborative modeling, e.g. interfaces for interconnecting sub-models, for an exchange of model components between different ALEM-T instances, and for distributed simulation capability. The modifications have to support the complete system development cycle: specification, infrastructure configuration, simulation, and evaluation. Further, it has to allow modeling of the complete model at different locations, at different times, and by different modelers or organizations. The latter issue means that the required collaborative modeling method has to be able to cope with different rules and processes as well as operational and organizational structures. In this sense, the methods should provide inter-organizational collaborative modeling [13]. The literature proposes several requirements and mechanisms collaborative modeling methods have to comply with, e.g. model viewers which display a model differently for each modeler [5, 13]. Further, model elements as well as the complete model have to be exportable into a common file format in order to present it at different places. A version control system is required, which allows comparison and backtracking of different model versions. A new user profile manager shall handle access rights of the users and a commentary function has to allow users to place asynchronously notes at the model workspace. However, most important is a mechanism that is able to split models into sub-models and to recombine them later [5]. In summary, a modeling method for autonomous logistic processes requires the following additional capabilities in order to circumnavigate problems emerging from organizational independence of modelers: – Spatial and temporal distributed modeling and simulation – Multi-user and multi-organization modeling
12
Scalability Effects in Modeling Autonomously Controlled Logistic Processes
165
A A
B B
ALEM Notation
ALEM Notation
Modeling
Concept Structure
Simulation
Procedure
Evaluation
Architecture
ToolSupport
Model
Simulation
Procedure
Evaluation
Architecture
Infrastructure
Company A
Legend
Modeling
Concept Structure
ToolSupport
Infrastructure
Company B
Interface
Information
Known Model Part
Unknown Model Part
Fig. 12.9 Multi-ALEM-framework [21]
– – – – – – –
Parallel and partial modeling of the logistic process and simulation run execution Capsulation of model elements and sub-models Standardized Interfaces between sub-models Exchange of sub-model an model information Exchange of simulation data Synchronization of model parts Quality assurance functionality (Syntactic, Consistency, Semantic)
Figure 12.9 presents exemplarily the exchange of sub-model information between different companies using the ALEM methodology. It shows the interfaces between both ALEM frameworks as well as between known and unknown sub-models. Company A knows six of nine sub-model elements, while company B knows only three of them. Both companies are either able to exchange information about the submodels content in order to learn more details of the overall model, or they exchange sub-model interface information. In the latter case, companies model their part of the logistic process independently. The interfaces, enabling interactions between different sub-models and ALEM frameworks’ processes, clearly have to state required and available information. While designing the overall process, each partner defines the information that is required for entering a process and that is provided on exit of a process. The sub-model interface hides other sub-models. The respective company notices known model elements as white boxes, while recognizing unknown model elements and their behavior as black boxes. Further, a new collaborative modeling view is required. This view has to be adjusted towards a specific users’ spatial and organizational location and represents the model in this context to him. Additionally, a new central model overview is required which shows the semantic and possible emergent behavior of decisions within sub-models. It reduces the modeling complexity of value networks.
166
B. Scholz-Reiter et al.
12.6 Summary and Outlook This article investigates challenges towards modeling autonomously controlled logistic systems, with regard to the scalability of the ALEM modeling methodology. It identifies two major sources of challenges: Challenges emerging from the increasing complexity of the systems and challenges originating from organizational independence between organizations participating in the autonomous processes. The complexity, induced by an increasing amount of objects, processes, and decision-strategies, makes it hard for logistic process experts to maintain a general view on all system aspects. Furthermore, with an increasing system size, modelers are hardly able to estimate the total effect of local modifications towards the overall system and towards a specific behavior in the overall system. Thus, the effort for refining and testing of a model increases. These difficulties occur in particular in case of miss-specified model elements or if error tracing is required. Moreover, an increasing number of objects, processes, and decisions prolong the time that is necessary for model refinement, error tracing, and simulation. The more complex a system is, the more time consuming are these tasks. Last but not least, logistic objects interact frequently with each other and react to random environmental events. Thus, results of simulation runs become more and more non-deterministic from an experts point of view and impede uncomplicated testing of a specific system behavior. Further challenges emerge from the organizational independence of companies involved in large value networks, like supply chains. Due to a logistic process expert’s limited view on partners’ or subcontractors’ processes, he might not be able to create a model covering the complete system in all details. Differences between partners’ control and information management strategies can prevent logistic process experts to obtain required information. For logistic process experts, this results in a chance either to neglect useful information, which could be obtained by the logistic objects, or to expect information, which is not passed or collected by a partner. In addition, the ALEM – Procedure Model assumes a central modeling of a highly decentralized system. It will work fine in case of a small logistic system that is under control of the modelers company. However, this approach might fail in case of larger systems, such as supply networks. The question is: How to model a decentralized autonomous system in a decentralized way across economically independent partners? Both issues picture different challenges of ALEM’s ability to scale up with a systems complexity as well as with its diversity in organizational manner. The paper presented first solution approaches for each challenge class that will be under investigation in future research.
12
Scalability Effects in Modeling Autonomously Controlled Logistic Processes
167
References 1. Allahverdi A, Ng C, Cheng T, Kovalyov M (2008) A survey of scheduling problems with setup times or costs. Eur J Oper Res 187(3):985–1032 2. Bauer B, Mueller J, Odell J (2001) Agent UML: a formalism for specifying multiagent software systems. Int J Software Eng Knowl Eng 11(3):207–230 3. Cooper MC, Lambert DM, Pagh JD (1997) Supply chain management: more than a new name for logistics. Int J Logist Manag 8(1):1–14 4. Dashkovskiy S, Wirth F, Jagalski T (2005) Autonomous control in shop floor logistics: analytic models. In: Chryssolouris G, Mourtzis D (eds) Manufacturing, modelling, management and control 2004. Elsevier, Amsterdam, NL 5. Gonzalez PP, Framiñan Torres JM (2008) Tools for collaborative business process modeling: classification and evaluation. Encyclopedia Netw Virtual Org 3:1643–1652 6. Hülsmann M, Windt K (2007) (eds) Understanding of autonomous cooperation and control in logistics – the impact of autonomy on management, information, communication and material flow. Springer, Berlin 7. Kolditz J (2009) Fachkonzeption für selbststeuernde logistische Prozesse. Ph.D. thesis, Universität Bremen, Berlin 8. Object Management Group: Unified Modeling Language (UML) V.2.2. Online. http://www. omg.org/technology/documents/modeling_spec_catalog.htm#UML 9. Odell J, Parunak H, Bauer B (2001) Representing agent interaction protocols in UML. In: Ciancarini P, Wooldridge M (eds) AOSE 2000, LNCS 1957. Springer, Berlin and Heidelberg, pp 121–140 10. Oestereich B, Weiss C, Schröder C, Weilkiens T, Lenhard A (2003) Objektorientierte Geschäftsprozessmodellierung mit der UML. Dpunkt, Heidelberg 11. Porter M (1985) Competitive advantage: creating and sustaining superior performance. Free Press, New York 12. Rekersbrink H, Makuschewitz T, Scholz-Reiter BA (2009) Distributed routing concept for vehicle routing problems. Logist Res 1(1):45–52 13. Ryu K, Yücesan E (2007) Cpm: a collaborative process modeling for cooperative manufacturers. Adv Eng Informat 21(2):231–239 14. Scholz-Reiter B, de Beer C, Freitag M, Jagalski T (2008) Bio-inspired and pheromone-based shop-floor control. Int J Comput Integrated Manuf 21(2):201–205 15. Scholz-Reiter B, Boese F, Jagalski T, Windt K (2007) Selbststeuerung in der betrieblichen Praxis: Ein Framework zur Auswahl der passenden Selbststeuerungsstrategie 16. Scholz-Reiter B, Görges M, Jagalski T, Mehrsai A (2009) Modelling and analysis of autonomously controlled production networks. In: Proceedings of the 13th IFAC Symposium on Information Control Problems in Manufacturing (INCOM 09). Moscow, Russia, pp 850–855 17. Scholz-Reiter B, Hildebrandt T, Kolditz J (2008) Modelling and simulation of autonomous logistic processes. In: Bruzzone A, Longo F, Piera M, Aguilar R, Frydman C (eds) 20th European modeling and simulation symposium. DIPTEM University of Genova, Amantea, Italy, pp 594–601 18. Scholz-Reiter B, Höhns, H (2006) Selbststeuerung logistischer Prozesse mit Agentensystemen. In: G. Schuh (ed) Produktionsplanung und -steuerung: Grundlagen, Gestaltung, Konzepte. Springer, Berlin, pp 745–780 19. Scholz-Reiter B, Höhns H, Kolditz J, Hildebrandt T (2005) Autonomous supply net coordination. In: Proceedings of 38th CIRP Manufacturing Systems Seminar. Florianopolis, Brazil. CD-ROM, 7 pages 20. Scholz-Reiter B, Kolditz J, Hildebrandt T (2009) Engineering autonomously controlled logistic systems. Int J Prod Res 47(6):1449–1468 21. Scholz-Reiter B, Rippel D, Sowade S (2011) Limitations in modeling autonomous logistic processes - challenges and solutions in business process modeling. In: IEEE International symposium on assembly and manufacturing (ISAM’11). Tampere, [in print]
168
B. Scholz-Reiter et al.
22. Scholz-Reiter B, Sowade S, Hildebrandt T, Rippel D (2009) Modelling of autonomous control explicitly considering orders as a kind of immaterial logistic objects in manufacturing processes. In: Proceedings of the 3rd international conference on changeable, agile, reconfigurable and virtual production (CARV 2009). München, Deutschland 23. Scholz-Reiter B, Sowade S, Hildebrandt T, Rippel D (2010) Modeling of orders in autonomously controlled logistic systems. Prod Eng Res Dev 4(4):319–325 24. Scholz-Reiter B, Sowade S, Rippel D, Teucke M, Özsahin M, Hildebrandt T (2009) A contribution to the application of autonomous control in manufacturing. Int J Comput 3(3):279–291 25. Scholz-Reiter B, Teucke M, Özsahin ME, Sowade S (2009) Smart label-supported autonomous supply chain control in the apparel industry. In: Lee Y, SaKong I (eds) Proceedings of the 5th international congress on logistics and SCM systems (ICLS2009). Overcoming economic crisis through collaborative logistics and supply chains, pp 44–52 26. Schreiber G, Akkermans H, Anjewierden A, Hoog R.d, Shadbolt N, Velde W.v.d, Wielinga B (2002) Knowledge engineering and management: the CommonKADS methodology. Artif Intell Med 24(1):97–102 27. Windt K (2008) Ermittlung des angemessenen selbststeuerungsgrades in der logistik – grenzen der selbststeuerung. In: Nyhuis P (ed) Beiträge zu einer theorie der logistik. Springer, Berlin, pp 349–372
Chapter 13
Exploitation of Manufacturing Flexibilities in Decision Methods for Autonomous Control of Production Processes: Findings from Industrial Practice and Theoretical Analysis Nicolas Gebhardt, Oliver Jeken, and Katja Windt
13.1 Introduction Nowadays series manufacturing traditionally shows a strict linkage of orders and products. The need to serve a high variance of customer requirements and to decrease internal production complexity at the same time provokes companies to standardize and simplify the processes of order management in production. Quick reactions on unpredicted disturbances (e.g. quality problems, production breakdowns, late order changes etc.) do include the reallocation of goods to customer orders rarely automatically but more as a manual action on emergencies. The fix linkage between orders and products lead to restricted possibilities in production control [25]. Turning this linkage into a flexible and dynamic factor in production planning and control is one of the potentials that autonomous control in production exploits. It aims at improving the achievement of logistic targets by decentralized, independent and autonomous decision processes in production planning and control. Autonomous control in production utilizes flexibilities inherent to the production system but mostly unused so far. One of these flexibilities encountered is the mentioned dynamic assignment of production parts to customer orders. This additional flexibility is called the flexibility of order allocation [25]. The main purpose of this contribution is to examine the observed practice in an industrial steel making case study as a good example of what autonomous control in production can look like. The case will be presented in the first chapter which explains the steel making production and the observed practice. The second chapter will then go on to present the theoretical background of autonomous production control, covering basic principle, main potential, and so far developed methods. Special emphasis is put on aspects relevant to the case study. Based on the definitions of autonomous control the third chapter inspects similarity of the N. Gebhardt, O. Jeken, and K. Windt (B) Global Production Logistics, International Logistics, School of Engineering and Science, Jacobs University Bremen, Bremen, Germany e-mail:
[email protected] M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_13,
169
170
N. Gebhardt et al.
phenomenon observed in industry and the principle of autonomous control. Finally conclusions are drawn from the case in line with comparability. In the last section a prospect is given of current and future matters of research regarding the research on autonomous control in production being on the cusp of first applications.
13.2 A Present Industrial Practice Akin to Autonomous Control in Production Even though research on applications of autonomous control methods for production control is still fairly infant, there are systems in current production giving an interesting insight of what applications of autonomous control can look like and what kind of limitations will appear. This chapter will introduce a case study of an industrial management practice comparable to autonomous control in production utilizing the flexibility of order allocation mentioned in the introduction. After a summary of theoretical perspective on autonomous control in the next chapter findings from the case are presented. The industry of steel making – especially sheet metal production – is characterized by a fairly linear production network and by highly variant products (Fig. 13.1). The main aspects of variety are the steel grade, dimensions, surface aspects and the type of material and coating. Make-to-order is the main strategy of order fulfillment. The single production steps do inherent complex sequencing rules depending on specific material and process parameters. This leads to a structure of production planning and control which exhibits centralized rough planning and order
blast furnance
steelplant (converter)
liquid steel
hot roller mill
continuous casting
cold roller mill
slap
annealing
anodization
hot coating
band coating
packaging
coil
photos from www.thyssenkrupp-technologies.com/de/presse/ and ArcelorMittal Bremen
Fig. 13.1 Process steps of steel sheet production (simplified chart)
13
Exploitation of Manufacturing Flexibilities in Decision Methods
171
management as well as decentralized short term planning and order sequencing of each production step [19]. Due to high product variety – the steel grade by itself offers several hundreds of options to choose from – and customer individual order properties the single customer orders are considered to be unique combinations of customer relevant attributes. Nevertheless, flexible handling of material consignment is a common practice in the industry of sheet steel production. Production steps are usually performed on consigned parts only but the order management can change the allocation of products to customer orders in between production steps. This additional flexibility in production originates from the need of quick reactions on schedule violations, quality issues etc. In such a case the order management can interchange the consignments of steel slaps or coils to improve over all order compliance. Triggering incidents of allocation changes come from different aspects. Reasons that appropriate an allocation change can be driven by: Aspects of the item (cast, slap or coil) and its current order allocation, e.g. quality
issues and rejects due to defects; they cause a violation of the order requirements by the item and the need to detach the item from the order. The production status of the currently allocated order, e.g. excess of safety surplus of material, order changes by the customer at short notice. The production statuses of all customer orders with special regard to deficient orders, e.g. shortfall quantities, scheduling delay, high prioritized orders. Figure 13.2 illustrates the process of allocation change. The affected objects and the information the allocation change is based on are highlighted. The main objective of performing order allocation changes is to increase the logistic targets achievement and order compliance. The logistics targets are short lead times, high due date reliability, high utilization, and low inventory [13]. In the present case study the main objective of material allocation changes is reduced to increasing order lateness. This is due to features of the production system which was examined. On the one hand non-time related logistic targets have not been the main focus and current necessity of order allocations. On the other hand they are as well not affected heavily enough by changes of item allocation to orders. Referring to the example shown in Fig. 13.2 the production item is processed while being consigned to the customer order “A”. At any time after the preceding process and before the start of the next pending process the consignment of the item can be changed. Customer order “B” might show shortfall in the number of steel coils or might be tied by a coil far behind the order production schedule. Any reason mentioned above concerning the properties of the former assigned order “A” or the item itself may interfere as well. In the example of Fig. 13.2 a saving in throughput time of the item can be achieved by swapping items. Considering the objective of lowering throughput times in the observed case the potential of time saving and thus processing an item faster is the main target of allocation changes. The examined practice of changing allocations of the production items from one customer order to another represents an action by the order management to reduce
172
N. Gebhardt et al. A
customer order A
item status
cold roller mill
hot roller mill allocation change B
customer order B
cold roller mill expected time saving time of allocation change
scheduled production step
steel coil as logistic object
time
start time order start time order B A
customer order
item properties
Fig. 13.2 Allocation change of production items to customer orders
interruptions in the value flow within production. This case gives a good example of what a practical application of autonomous control in production can look like. The observed practice utilizes the additional flexibility of order allocation in a way that exhibits many of the characteristics of autonomous control. The total number of allocation changes per production output is large enough to draw significant conclusions1. For the research of autonomous control in production two interesting question arise: First, to what extent does the practice of allocation changes in the steel industry
comply with the idea and definitions of autonomous control in production? Second, do the findings of allocation changes in the steel industry increase the
level of achievement of the logistic objectives in production? These questions are attended in the chapter “Comparison of the order allocation practice in steel making and autonomous control” below. First, the following chapter presents the theory of autonomous control in production and autonomous control methods.
1
Due to reasons of secrecy obligation of the observed production company more details cannot be exposed.
13
Exploitation of Manufacturing Flexibilities in Decision Methods
173
13.3 Theoretical Perspective of Autonomous Control in Production 13.3.1 Basic Principle of Autonomous Control in Production The Collaborative Research Centre 637 “Autonomous Cooperating Logistic Processes: A Paradigm Shift and its Limitations” at University of Bremen and Jacobs University Bremen underlies a main hypothesis to autonomous control: A higher level of decentralized and heterarchical decision making can result in a better achievement of logistic objectives compared to conventionally managed processes despite increasing complexity [27]. Conventional methods of production planning and control manage complexity in manufacturing by mostly inflexible scheduling and centralized computing processes. Regarding all available information in a centralized scheduling task will often lead to a very complex calculation. Schedule optimization usually leads to NP-hard problems where solution space grows faster than the speed of decision-making. Due to this fact central heuristics are used which cannot achieve optimal solutions or react on quick changes. The idea of autonomous control in production tries to break this tie. Decisions are taken on the level of single parts and only regarding confined spaces of information and consequences. The decision capabilities are shifted from a central control to the system elements [9]. Distributed heuristics that operate on local knowledge can result in low decision complexity and acceptable performance. Single decisions are taken independently of each other on the basis of only a fraction of the global data. The range of impact is limited to only a few of the production steps pending next and to only one logistic object. Thus the logic functions of decision making become fairly simple – compared to centralized control methods – and can be executed much faster. The decentralization and the independence of decision making are main criteria of autonomous control. The processes of decision making are performed by the logistic objects or their agents [15]. Autonomous control in the field of logistics is defined as “. . . the ability of logistics objects to process information, to render and to execute decisions on their own. (. . . ) Logistic objects (e.g. part, pallet, order, or work station) that are able to fulfill these conditions are called intelligent objects.” [27]
The idea includes the assumption that this strategy of decentralized and at first glance egoistic decision making will have a beneficial impact on overall logistic objectives of the whole system. Due to distributed and consequently more flexible decisions autonomous control can react better on high dynamics and complexity and thus increase system robustness [8]. To make this possible the deciding functions must regard not only their own state and the state of surrounding logistic objects like production systems but also the basic figures of an incorporated parameter set of business objectives. Furthermore the current aim of the production system – i.e. the production program or order book – must be part of the information the decisions are based on. The concept of autonomous control in production is in line with such approaches as “intelligent
174
N. Gebhardt et al.
products” [12] and “product-driven control” [6], but extends these by a loose link of products and orders.
13.3.2 Manufacturing Flexibilities as the Main Potential of Autonomous Control in Production A conventional production system will not provide decision alternatives in the meaning and suitable extend of autonomous control in production. However, the number of possible decision nodes and decision alternatives is crucial for the efficiency of autonomous control in logistics. The more decision nodes with a high number of decision alternatives exist within a logistic system, the higher is the logistics potential that can be realized with autonomous control methods. Alternative options in regard to the decision capacity of autonomous control in production systems can be provided by the degree of manufacturing flexibility as it offers multiple ways to perform a manufacturing process [25]. Manufacturing flexibility is generally regarded as the ability to adapt. Definitions related to manufacturing flexibility follow the idea of adaptability of a manufacturing system to uncertainties [4, 11]. “Manufacturing flexibility is (. . . ) about alternatives that suit certain conditions from the outset.” [25]
In order to implement autonomous control in production logistics the vision of an intelligent logistic object (e.g. part) or its representing agent is the underlying scenario. From the point of view of a single logistic object within a production system possible decisions can be categorized into selections of:
The allocation to a suiting customer order The product variant within the remaining scope The resources to be processed by The sequence of manufacturing steps
The research results of the CRC 637 include a catalogue of manufacturing flexibility types that represents the decision space of autonomous control in production. By this it will even be possible to derive a set of indicators for the degree of autonomous control flexibility provided by the production system. The catalogue of five types of manufacturing flexibilities was established in a systematical structure incorporating all common elements constituting a manufacturing system. The catalogue represents a set of independent measures for manufacturing flexibilities useable in autonomous control in production as a provider for decision alternatives [25]. Table 13.1 shows the catalogue of manufacturing flexibility types including degrees of freedom which can be formulated to operationalize manufacturing flexibility. From the orders as logical element a new type of manufacturing flexibility arises, called allocation flexibility. It describes the flexibility of an order to be allocated to a different product or its unfinished precursor during a manufacturing process. An order specifies one or more product variants and due dates for the production.
13
Exploitation of Manufacturing Flexibilities in Decision Methods
175
Table 13.1 Catalogue of manufacturing flexibilities [25] Level
Element
Flexibility Type
logical
order
Allocation Flexibility
convertible orders
Machine Flexibility
different operations
resource
Degree of Freedom
Material handling Flexibility
multiple system paths
Volume Flexibility
workload variation
A B C
physical
product
Operation Flexibility
different processing plans
A product variant is a set of features where some features have slightly different specifications compared to the set of features of another product [7]. Production resources and products are on the physical level of the manufacturing system. Related to the element resource are machine flexibility, material handling flexibility and volume flexibility. The element product refers to operation flexibility. The design of the product directly causes a certain limitation in possible alternatives of production steps. Process planning and scheduling deduce production plans from the product data and thereby form a critical stage in the exploitation of manufacturing flexibilities for autonomous control in production.
13.3.3 Further Details of the Flexibility of Order Allocation In general these manufacturing flexibility types are not unknown but fairly well researched, listed and classified [20]. Nevertheless the allocation flexibility of products to orders is so far rarely used but an important potential to autonomous control in production. “Allocation flexibility is related to the order as it describes the flexibility of an order to be allocated to a different product or its unfinished precursor.” [25]
Allocation flexibility depends on the availability of orders that a certain product can be allocated to. That means that a product, component or part in an early production state will most likely find a higher amount of suiting orders since it is less specified
176
N. Gebhardt et al. Limitation of alternatives due to production progress
Extension due to additional customer order Final product alternative
Start of Manufacturing Cycle
End of Manufacturing Cycle
Possible decision alternatives due to combination of production process steps, machines, final product variants and customer orders
Fig. 13.3 The Product Variant Corridor as a representation of allocation flexibility (Windt and Jeken 2009)
by further production processes. The amount of orders that a product can be allocated to at a certain point of time does depend on static and dynamic information. The product structure and the production system layout are static whereas the status of the order book and the features of the regarded part are dynamic. Altogether these types of information define the decision alternatives concerning order allocation. Supporting the utilization of allocation flexibility for autonomous control in production the variant corridor has been developed (cf. Fig. 13.3). The variant corridor is an approach providing each item with all combination of product variant and customer orders it can currently be allocated to. In short, the variant corridor assembles all decision alternatives of a part or component for each manufacturing step. It then draws off any alternative not obtainable anymore due to an advanced process stage of the product or not included in the current orders. Thus the product variant corridor represents the range of possible production alternatives for a given production stage and matches them with the actual customer orders. Each production step applies more features to the product parts or components. A feature is a property that is added to a product (geometries, mechanical behavior, mounting etc.). Thus a production step can only decrease the amount variants possible to be met by the product, part or component and the corridor narrows down along production progress. However, it can expand again in the case of new emerging and obtainable customer order. Only after a customer decoupling point, the potential a dynamic product-order allocation is exhausted [7].
13.3.4 Methods of Autonomous Control in Production The principle of autonomous control in production leads to two separate fields of prerequisites of practical application. First logistic objects need to have the technological capability to be identified and traced within the production system and to share information with the system environment. Secondly the fully decentralized scheme of decision making calls for the technical and methodical capability
13
Exploitation of Manufacturing Flexibilities in Decision Methods
177
of processing data and rendering decision. This chapter discusses the methodical requirements of autonomous Control for logistic items in production systems. The actual execution of logistic decisions requires the capability of the production system to act upon the items’ commands. In order to execute a decision alternative the product has to communicate the decision to the material flow system, which then organizes the necessary processes [7]. The ability of logistic items to autonomously navigate through a production system becomes more and more possible since recent developments (e.g. information and communication technologies, e.g. radio frequency identification, global positioning system or universal mobile telecommunications system) [3]. Studies on computing times of the decision making process show that autonomous control methods derive decisions faster than conventional approaches can do [28]. Nevertheless a boundary of autonomous control that needs to be dealt with in future will be the calculating capacity of logistic objects. Running times of single decisions in autonomously controlled systems increase with the system complexity and level of autonomous control. These two aspects in combination can lead to decision times too long for the dynamics of the certain system. The question is how much computing capacity can be fitted under economically reasonable conditions. Autonomous control might suffer in the possible level of computing capacity due to practical and economical reasons. In an autonomously controlled system each system element has to be a decision unit equipped with decision-making competence according to the current task [5]. Concerning the available options provided by the manufacturing flexibilities shown in Table 13.1 the decision process of a logistic item taking autonomous control has to fulfill the following set of tasks. According to decision theory the process of decision making incorporates five sub processes [10] which are:
Describing the problem Defining the target system Generating decision alternatives Evaluating the decision alternatives according to target system and Executing the decision alternative with the best target contribution
Concerning a typical job-shop manufacturing the logistic items have to be able to decide about the next production process step, according to which product variant it decides for, on which machine and for which customer order it will get manufactured. The decision space derived by these tasks is shown in Fig 13.4 Within the decision-making, gathering the decision alternatives is critical to the performance of autonomous control in production. As described above each decision is taken by or for a single item and before each production step. The decision alternatives are especially available for product structures with many variants, as one component of a product can precede for different final products variants and for different customers. Generating decision alternatives requires the necessary information from the relevant environment of the item. In addition to generating the alternatives there is a need for further information in order to evaluate these alternatives. All this information encompasses:
178
N. Gebhardt et al.
Fig. 13.4 Decision-making of autonomously controlling logistic items in a typical job-shop manufacturing [24]
The amount of obtainable final product variants The possible production processes The accessible machines for the different available production steps The current situation at the available machines (capacity, work in progress, planed idle times, machine breakdowns etc.) The current demand situation In the research work of the CRC 637 several different autonomous control methods have been developed and tested. Others were invented without explicitly naming them autonomous control methods and have been integrated into the approaches of the CRC 637. Several studies comparing the performance and behavior of autonomous control methods have illustrated that autonomous control can realize a higher logistics target achievement in comparison to conventional production planning and control [14, 17, 23]. “Autonomous control methods describe the logistic objects’ target systems, the way they interact and how and when decisions are taken.” [26] “An autonomous control method is a generic algorithm that describes how logistics objects render and execute decisions by their own.” [23]
A key demand on autonomous control methods is the lack of influencing the basic functionality of the production process in terms of process structure, elements and tasks. There are many different ways an autonomous control method can operate. Always it will have to be designed or at least adjusted to an existing logistic process. In this volume, the article “A Comparative View on Existing Autonomous Control Approaches” presents simulation results comparing the performance of several
13
Exploitation of Manufacturing Flexibilities in Decision Methods
179
autonomous control algorithms. In [22] and [23] one can find simulation studies analyzing the behavior of the most advanced and promising methods so far. The table below lists examples of autonomous control methods along with their principles. The “Ant Pheromone” method for example is inspired by ants’ foraging behavior. It uses virtual pheromones emitted by the parts at a station when they are processed by it; pheromones add up with the already existing amounts at the station; over time the pheromones slowly evaporate. Stations with higher pheromone levels have better throughput and are therefore preferred by following items. The “Simple Rule Based” method represents an example from a group of methods. In this method each semi-finished part chooses the next production step by preferring the machine with the lowest number of waiting items in front of it. Autonomous control methods showing similar attributes in design and principles can be pooled in an autonomous control strategy. “An autonomous control strategy is a generic term that summarizes multiple autonomous control methods which have common design patterns.” [26]
Examples for autonomous control strategies are “Rational Strategies” which use statistical approaches in order to predict future states of the logistic system for taking decisions [16]. This could be calculating the past average throughput time or queue lengths for each process step. “Bounded Rationality” is an autonomous control strategy deciding with help of simple heuristics without performing large data collection or intensive calculation [18]. The category of “Combined Strategies” summarizes approaches combing autonomous control methods from the other categories in parallel and using a weighted average of the different decision results [16]. The decision function of a product in an autonomously controlled production system will operate in a decision space based of the manufacturing flexibilities described above (refer to Table 13.1 and Fig. 13.4). Consequently, the physical results of a single decision will be the next production step of an item and the allocated machine or other kind of logistic resource. Informational results are the allocation of the item to a certain product variant as production outcome and to a certain customer order [24]. Decision criteria on the other hand can be derived from the available information listed above. The decision criteria taken into account incorporate information from local and global domains of the production system. Each system element in an autonomously controlled system is characterized by target oriented behavior. This means that global objectives – e.g. provided by the corporate management – can be modified independently by the decision functions in compliance with the objectives of the autonomous item. The decision functions of a specific autonomous control method can derive an individual set of weighting coefficients for each item and decision, consequently blending the strategic positioning of a company and the ideal decision from the items’ perspective [21]. Table 13.2 shows a selection of autonomous control methods developed and tested so far. Studies on the topic of autonomous control methods classification show that the existing methods mainly differ from each other in terms of performance and
180
N. Gebhardt et al.
Table 13.2 Selected examples for autonomous control methods from a simulation approach on a car terminal [23] Name Description Ant pheromone After an item is processed at a work station it leaves a fixed amount of pheromones which add up with the already existing amounts at the station; over time the pheromones slowly evaporate Stations with higher pheromone levels are stations with better throughput and are therefore preferred by the following items DLRP Parts request a route from machines; machines communicate best routes to a destination Holonic Two agents bargain over the next item to be processed. Agents are machines and management; the management punishes machines for delays and machines bid to get jobs from management Queue length estimation Choose work station with lowest queue length (number of items) Simple rule based Compares estimated waiting time at buffers using future events. The parts are rated in estimated processing times and current buffer levels are calculated as a sum of them. Parts choose the machine with the lowest processing time buffer
applicability. The most important distinctions have been discovered in the following criteria: Temporal data can be used by the method from the past, future, or both. Methods exhibit different number of planning steps, i.e. the number of considered
production steps in the future. The decision process can create artificial values (apart from external informa-
tion), e.g. pheromones or virtual costs, and then use or pass this data. The type of communication of items, machines and data bases differ. Methods will use various data scopes for decision making. The decision itself can be calculated by different actors, e.g. the items, machines,
agents etc., and The place of data storage can alter.
These criteria form a set of parameters eligible to classify autonomous control methods [23]. In this volume, the article “A Comparative View on Existing Autonomous Control Approaches” simulation results are presented which show the differences in performance of several autonomous control methods in an industrial application. The study shows that autonomous control methods will have to be carefully chosen, adapted or even completely designed for a specific application scenario. Tailor-made solutions will be the first way of implementation for a considerable long time. Autonomous control in production rests upon flexibilities of a production system. These flexibilities have been limited or left unused so far for the sake of simplifying the centralized production planning and control. The decentralized and autonomous decision strategy of autonomous control is restricted to local precincts. This enables autonomous control to exploit the inherent flexibilities in order to improve the system performance.
13
Exploitation of Manufacturing Flexibilities in Decision Methods
181
13.4 Comparison of the Order Allocation Practice in Steel Making and Autonomous Control in Production From the case study presented in the above chapter “A present industrial practice akin to autonomous control in production” conclusions can be drawn concerning autonomous control. In this chapter the question of comparability and generalizability of the case will be dealt with (as stated at the beginning of this article). After that the findings and conclusions are presented. In order to find to what extent these conclusions are valid for autonomous control first the compliance of the observed phenomenon with the idea and definitions of autonomous control in production has to be identified. If there is a certain scope of compliance the findings about the phenomenon can be generalized to autonomous control within the extent of this scope [2]. The performance of allocation changes in the steel industry can give an indication of the performance of autonomous control if they are proven to be similar.
13.4.1 Similarity of Observed Phenomenon in Steel Industry to Autonomous Control The similarity of the observed practice and autonomous control is tested by applying the main definitions of autonomous control to the case. The definitions are presented in the chapter “Theoretical perspective of autonomous control in production” above. The resulting test questions are listed in Table 13.3. The findings of the assessment questions are discussed below.
Table 13.3 Test of similarity between the observed allocation practice in steel industry and autonomous control No. Subject Testing question Finding 1 Exploited Is the exploited flexibility within scope of Yes flexibility autonomous control in production? 2 Autonomy Do the logistics objects process information, render No and execute decisions by themselves? 3 Structure Is the planning and controlling decentralized and Mostly not heterarchical? 4 Impact Is the range of impact limited to only the production Yes step pending next? 5 Is the range of impact limited to only one logistic Mostly yes object? 6 Information Are decisions taken only regarding confined spaces Yes of information? 7 Decision making Decisions are taken independently of each other? Mostly yes 8 Are the functions of decision making are fairly Yes simple compared to a centralized strategy?
182
N. Gebhardt et al.
Question number one in Table 13.3 refers to the manufacturing flexibilities which are seen as the underlying potential which can be utilized by autonomous control in production. However, a positive answer to this question is not mandatory to prove similarity of the case and autonomous control since finding a new flexibility cannot be foreclosed. In this case the exploited flexibility matches the idea and definition of the allocation flexibility described above (refer to chapter “Further details of the flexibility of order allocation”). For the new allocation of items only the shortfall quantities of existing manufacturing orders are taken into account. The second question aims at one of the main objectives of autonomous control in production which is the shift of planning and scheduling tasks to the logistic items. Of course this is not the case in the observed phenomenon. Hence no conclusions will be valid related to the physical implementation of decision making. The question is, though, if the discovered processes of decision making equals to the idea of autonomous control. The following checks number three to eight consider this aspect and therefore answer this question. The process of production planning and controlling is not decentralized. However, changing the allocation of items can be performed by the local control of workstations. Questions number four and five deal with the effects on the production throughput that a single decision of an allocation change have. The parts in production are consigned to a certain order. Changes are not taken back after the upcoming production step and the new allocation will remain until the next change or finishing of production. The production sequences only depend on the specific customer order. Concerning the allocation change itself the decision does only affect the production plan of a single item. Nevertheless, if the reasons leading to an allocation change described above apply to more than one item, the conditions allow considering this as an accumulation of single events. The question number six in Table 13.3 addresses information processing. In autonomous control only a limited amount of information shall be taken into account in each decision process. By this the decision complexity remains low and the decision speed can be maintained. In the observed case the space of information consists of the current production state of the item, the attributes of the order it is allocated to and the orders showing any applicable deficiency. Concerning the decision function of allocation change the independence and the simplicity are to be assessed. The characteristics of the observed practice cover a broad range of these attributes. Compared to a complete and central optimization of all items’ allocation the decision making is fairly independent and simple. The comparison of the allocation practice in the steel industry and autonomous control shows that findings from the case study can be transferred regarding most aspects. Conclusions about facts based on the independence of decisions are to be seen invalid in the first instance, as well as conclusions about the technological application of autonomous control. Certainly the attributes of independence and decentralization are vital aspects of autonomous control. As a result this case study can give a good insight and incitement to the development of autonomous control but cannot achieve the preconditions of an example application.
13
Exploitation of Manufacturing Flexibilities in Decision Methods
183
13.4.2 Observed Benefits of Dynamic Order Allocation Practice in Steel Making As described above the main intent of the allocation changes in steel industry is to improve due date reliability and order compliance in production. Non-time related logistic targets are not affected heavily enough by changes of item allocation to orders. The following Fig. 13.5 shows a summary of the findings about the order allocation changes. The study includes the five process steps with the highest number of order allocation changes as illustrated in Fig. 13.5. Altogether, theses five production steps raised nearly all of the allocation changes.
CC & AD steelplant (converter)
blast furnance
CC & AD:
fracion of over all allocation changes
HR: HC: BC:
continuous casting
continuous casting & adjusting hot rolling hot coating band coating
HR hot roller mill
cold roller mill
annealing
anodization
hot coating
band coating
HC
BC
packaging
70%
60% 50%
production steps
40% 30% 20%
10% 0%
CC
AD steel making and casting
expected time saving in the moment of allocation change
HR
HC
BC
rolling and coating processes
Ø 7,3 ± s 3,2 days
Ø 0 ± s 2,1 days
Ø 77 % * Ø 98 % *
Ø 161 % *
output lateness of loosing orders obtaining orders
Ø 103 % *
* du to data protection the output lateness is shown as percentage of the average lateness of all customer orders at delivery
Fig. 13.5 Figures of allocation change performance
184
N. Gebhardt et al.
Most of the allocation changes happen in earlier production steps. The high product variety of the steel coil industry shifts a high number of allocation changes in the very beginning of the production where many product variances can still be fabricated from each single item. Each production distinguishes roughly the same portion of final product features. Thus the variant corridor of the steel production will constantly shrink along the production process (refer to Fig. 13.3). New customer orders are released by introducing new items into the production, thereby leaving no shortfall quantities right away which could be assigned by the production items along the process. Due to reasons of secrecy obligation in this case the total number of allocation changes per production output cannot be released. However, the amount of events is large enough to draw significant conclusions. Figures show that together with the diminishing amount of decision alternatives the number of allocation changes abates quickly along the production sequence. At late production steps less time remains to meet the customer orders in terms of quality and quantity. A smaller number of decision alternatives give less potential of improvements by reallocating the product items. The potential of order allocation changes becomes evident in the expected time saving measured as shown in Fig. 13.2. The decision of the new allocation is primarily based on a comparison of the production schedules of the considerable orders. The aim is to process the items as fast as possible by allocating them to the most urgently pending one among the suitable orders. The allocation changes performed in the early production steps of steel making feature much higher time savings. The allocation changes conducted in late production steps hardly show any time saving at all. Apart from the fact that less decision alternatives can be expected and exploited in late production, these steps exhibit a lower variance of throughput times – thus lessening the need for compensation of lateness by allocation changes. Finally, in Fig. 13.5 the lateness of all orders affected by the allocation changes in the certain production levels are shown. Interestingly enough, orders that where influenced by an allocation change in early production steps were finished on time or even early. Orders affected by allocation changes at the end of the production sequence, however, show a significant lateness. These findings correspond to the observed expected time savings and to the industry experts’ anticipations on the topic. The observed practice of order allocation strategy similar to autonomous control is qualified as a good perspective on autonomous control in production. Dynamic and decentralized reallocation of production items is successfully used in the steel industry for improving due date reliability. Nevertheless, the later the changes in allocations are performed, the less the effect on the objectives. As a possible conclusion for the research of autonomous control in production the findings indicate boundary conditions for the exploitation of allocation flexibility as a potential for autonomous control. Furthermore the case shows the applicability of the catalogue of manufacturing flexibilities (refer to Fig. 13.4) as a basic set of measures for autonomous control potentials.
13
Exploitation of Manufacturing Flexibilities in Decision Methods
185
13.5 Prospects on Current Research of Autonomously Controlled Logistic Systems Research work of autonomous control in production is currently passing the cusp of application. Several studies show that autonomous control utilizing flexibilities in logistic systems as new decision alternatives can improve the achievement of logistic targets. At the same time simulations have been conducted that display the divergent properties and behavior of methods of autonomous control in production. This raises the necessity of focusing future research on the limitations and characteristics of single methods of autonomous control in production since their diversity will place new dimensions in the research of autonomous control. They all will have to be analyzed with their own impact on the limitations of autonomous control. The contribution in this volume “A Comparative View on Existing Autonomous Control Approaches” represents a first access into this field of research. The presented case of dynamic order allocation practice in industry gives a good example of what an application of autonomous control in production can look like. The case emphasizes the potential of autonomous solutions that act on sub processes of production or sub tasks in planning and control. At the same time the case demonstrates that autonomous control in production strongly depends on manufacturing flexibilities as elementary prerequisites. Flexibilities that can be exploited by autonomous control can be classified by the established catalogue of manufacturing flexibilities. In order to design an economically reasonable application of autonomous control in production these flexibilities will have to be identified in the specific system. They are determined not only by the production system but by the product structure and design as well as they majorly restrict process design in production and thus the potentials of autonomous control. In future research enhancing the performance of autonomously controlled production will have to be focused on the collaborative identification of manufacturing flexibilities that can support autonomous control performance. Autonomous control solutions coming from an integrated perspective on the production resources, products, planning and control are most promising for offering the high potentials of increasing the achievement of logistic targets in production by autonomous control.
References 1. Armbruster D, de Beer C, Freitag M, Jagalski T, Ringhofer C (2006) Autonomous control of production networks using a pheromone approach. Phys A 363(1):104–114 2. Atteslander P (2003) Methoden der empirischen sozialforschung. de Gruyter, Berlin 3. Böse F, Lampe W (2005) Adoption of RFID in logistics. In: Proceedings of IBIMA international business information management association conference, Cairo 4. ElMaraghy HA (2006) Flexible and reconfigurable manufacturing systems paradigms. Int J Flexible Manuf Syst 17(4):261–276
186
N. Gebhardt et al.
5. Frese E, Schmidt G, Hahn D, Horvath P (1996) Organisationsstrukturen und Management. In: Eversheim W, Schuh G (eds) Betriebshütte, produktion und management, 7th edn. Springer, Berlin 6. Gouyon D, David M (2008) Implementing the concept of product-driven control using wireless sensor networks: some experiments and issues. In: 17th IFAC World Congress 17(1) 7. Hribernik K, Pille C, Jeken O, Thoben K-D, Windt K, Busse M (2010) Autonomous control of intelligent products in beginning of life processes. In: Proceedings of the 7th international conference on product lifecycle management 8. Hülsmann M, Windt K (2007) Understanding autonomous cooperation and control in logistics – the impact on management, information and communication and material flow. Springer, Berlin 9. Krallmann H (2004) Systemanalyse in Unternehmen: partizipative Vorgehensmodelle, objektund prozessorientierte Analysen, flexible Organisationsarchitekturen, 3rd edn. Oldenbourg, München, Wien 10. Laux H (2007) Entscheidungstheorie, 7th edn. Springer, Berlin 11. Mandelbaum M (1978) Flexibility in decision making: an exploration and unification 12. McFarlane D, Sarma S, Lung Chirn J, Wong C, Ashton K (2003) Auto ID systems and intelligent manufacturing control. Eng Appl Artif Intell 16(4):365–376 13. Nyhuis P, Wiendahl H (2008) Fundamentals of production logistics: theory, tools and applications. Springer, New York 14. Rekersbrink H, Makuschewitz T, Scholz-Reiter B (2009) A distributed routing concept for vehicle routing problems. Logist Res 1(1):45–52 15. Scholz-Reiter B, Windt K, Kolditz J, Boese F, Hildebrandt T, Philipp T, et al. (2004) New concepts of modeling and evaluating autonomous logistic processes. In: Proceedings of IFAC conference on manufacturing, modeling, management and control 16. Scholz-Reiter B, Jagalski T, Bendul J (2007) Autonomous control of a shop floor based on bee’s foraging behavior. In: Haasis H-D, Kreowski H-J, Scholz-Reiter B (eds) First international conference on dynamics in logistics (LDIC 2007). 514–423. Springer, Berlin 17. Scholz-Reiter B, Görges M, Philipp T (2009) Autonomously controlled production systems – influence of autonomous control level on logistic performance. CIRP Ann Manuf Tech 58(1):395–398 18. Simon HA (1997) Models of bounded rationality. MIT, Cambridge, MA 19. Tang L, Liu J, Rong A, Yang Z (2001) A review of planning and scheduling systems and methods for integrated steel production, Eur J Oper Res 133(1):1–20 20. Wiendahl H-P, ElMaraghy HA, Nyhuis P, Zäh MF, Wiendahl H-H, Duffie N, Brieke M (2007) Changeable manufacturing - classification, design and operation. CIRP Ann Manuf Tech 56(2):783–809 21. Windt K, Jeken O, Arbabzadah F (2010) Improved logistics performance through the use of locked flexibility potentials. In: 43rd CIRP international conference on manufacturing systems May 2010. 892–899 22. Windt K, Becker T Jeken O, Gelessus A (2010) A classification pattern for autonomous control methods in logistics. Logist Res 2(2):109–120 23. Windt K, Becker T, Kolev I (2010) A comparison of the logistics performance of autonomous control methods in production logistics. In: Sihn W, Kuhlang P (eds) Proceedings of the 43rd CIRP international conference on manufacturing systems. Sustainable Production and Logistics in Global Networks, 576–583. Springer, New York 24. Windt K, Jeken O, Gebhardt N (2010) Nutzung und Erweiterung von Flexibilitätspotentialen in Fertigung und Montage durch Selbststeuerung und Konstruktion, 23. HAB-Forschungsseminar “Wandlungsfähige Produktionssysteme”, Hannover, Germany, October 2010 25. Windt K, Jeken O (2009) Allocation flexibility - a new flexibility type as an enabler for autonomous control in production logistics. In: 42nd CIRP conference on manufacturing systems 26. Windt K, Becker T (2009) Applying autonomous control methods in different logistic processes - a comparison by using an autonomous control application matrix. In; Proceedings of the 17th mediterranean conference on control and automation, 1451–1455.
13
Exploitation of Manufacturing Flexibilities in Decision Methods
187
27. Windt K, Böse F, Philipp T (2008) Autonomy in production logistics – identification, characterisation and application. Int J Robot CIM 24(4):572–578 28. Windt K (2008) Ermittlung des angemessenen Selbststeuerungsgrades in der Logistik – Grenzen der Selbststeuerung. In: Nyhuis P (ed) Beiträge zu einer Theorie der Logistik, pp 349–372. Springer, Berlin
Part III
Technological Contributions and Limitations
Chapter 14
Views on Technological Contributions and Limitations of Autonomous Cooperation and Control in Logistics Jakub Piotrowski and Bernd Scholz-Reiter
Today, planning and control of logistic processes is generally executed by centralized logistic systems which are not able to cope with the high requirements for flexible order processing due to increasing dynamics and complexity. In many cases exiting planning and control systems do not allow for a fast and flexible adaptation to changing environmental influences. This weakness is often caused by a non-synchronized information and material flow in logistic systems. But there is an ongoing paradigm shift from centralized control of logistic processes in hierarchical structures towards decentralized control in heterarchical structures. The concept of autonomously controlled logistic systems as an approach of a decentralised planning and control system require “intelligent” logistic objects. Autonomy in this case means, that the logistic system, the processes and objects have the capability to design their input-, throughput- and output-profiles as an anticipative or reactive answer to changing environmental parameters. Logistic objects in autonomously controlled logistics systems would operate independently according to their own objectives by processing information, make decisions and executing the decisions by themselves [1]. The aim of autonomous control mechanisms is to increase robustness and positive emergence of logistics systems by a distributed and flexible coping with dynamics and complexity [2]. In several simulation studies the use of autonomous control causes more flexibility, adaptivity and robustness to logistics systems. This is the case especially in logistic scenarios with a high degree of dynamics and complexity, for example caused by express orders, machine failures [3] or the decreasing remaining shelf life time of food products caused by unexpected situations during the logistic processes. The development of information and communication technologies, e.g. the identification technologies like RFID (Radio Frequency Identification), communication technologies like GPRS (General Packet Radio Service) or UMTS (Universal
J. Piotrowski (B) and B. Scholz-Reiter Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany e-mail:
[email protected] M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_14,
191
192
J. Piotrowski and B. Scholz-Reiter
Mobile Telecommunications System), makes “intelligent” logistic objects (and therefore autonomously controlled logistics systems) possible. Decentralized decision-making with autonomous control methods requires a real-time information flow to enable logistic objects to analyse the current system situation and use this data to decide which subsequent steps have to be taken. To realise this idea, identification technologies like RFID, real-time locating systems like GPS (Global Positioning System) and communication technologies like GSM (Global System for Mobile Communications) or GPRS and UMTS are necessary which continuously become smaller and cheaper. On the one hand, logistic objects are able to collect information about their own situation. On the other hand, this information can be passed through to other logistic objects if required. In the real world there are many difficulties to realise autonomous control or autonomously cooperating logistic objects. In most cases, logistic objects are not able to communicate, they have neither sensors that can be used to collect state data nor processing modules that can analyse these data and execute the decisions by themselves. Imagine a package that wants to route through a logistic system autonomously. There is a couple of data required to realise the idea of autonomous control. There is also a need of information about available route infrastructure, available trucks at any transfer point or information about traffic situation on routes. Furthermore, the logistic objects have to cooperate with each other in a highly dynamic and heterogeneous logistic system. This cooperation requires the integration of these objects as data sources and also as data collecting objects. Today, most of the logistic objects are not able to collect these decision relevant data by themselves. In order to enable this, the use of additional information and communication technologies is necessary. In many cases, the representation of these objects is realised by using multi agent systems (MAS) to enable a self decisionmaking of non intelligent logistic objects [4, 5]. The authors of “Implications of Communication Constraints for the DLRP in Transport Logistics”, Bernd-Ludwig Wenning, Christian Zabel, Henning Rekersbrink, Carmelita Görg and Bernd Scholz-Reiter, aim at the investigation of distributed routing algorithms transport processes and the technical constraints such as limited communication bandwidth or storage capacity and indirect constraints that are caused, for example, by the communication costs. The idea of distributed routing algorithms is already successful in communication networks. The concept of distributed routing can be assigned into logistic networks as a Distributed Logistic Routing Protocol (DLRP). The authors discuss the mentioned constraints and their implications for the implementation and operation of DLRP in transport logistic processes in their contribution and focus on communication volume. The chapter “Embedded Intelligent Objects in Food Logistics – Technical Limits of Local Decision Making” by Reiner Jedermann, Javier Palafox-Albarran, Amir Jabarri, and Walter Lang deals on the one hand with wireless sensors as a platform for local decision making. On the other hand, the authors discuss networks of wireless sensors as a data source for autonomous logistic processes. A major challenge is the reliability and range of communication as well as the required energy resource.
14
Views on Technological Contributions and Limitations
193
Autonomous control methods themselves, like algorithms for local data evaluation, can help to overcome these first restrictions during the implementations using the example of food logistics. However, the realisation of the paradigm of autonomous control by giving the “intelligence” directly to the logistic object is often not reasonable from the economic point of view. As mentioned above, it is possible to represent logistic objects in a virtual environment like MAS to enable a decentralised decision-making of these objects. Tobias Warden, Robert Porzel, Jan D. Gehrke, Hagen Langer, Otthein Herzog, and Rainer Malaka introduce in “Knowledge Management for Agent-based Control under Temporal Bounds” how agents, as a medium to realize autonomous cooperating logistic objects, can create tailored models which support their respective decision making. These models and knowledge management tasks, such as the adaption of existing and the compilation of new models, therefore need to be performed in a suitable time. Furthermore, the authors discuss how the agents can maintain their local models continuously while at the same time adhering to their logistic objectives. Another task within the technical implementation of autonomous logistics is to synchronise real-world data with the data within the MAS. The task of data integration has a significant position especially in dynamic logistic systems with a high flow of objects and crucial data for decision making in terms of autonomous control. The chapter “Impacts of Data Integration Approaches on the Limitations of Autonomous Cooperating Logistics Processes” contributed by Karl A. Hribernik, Christoph Kramer, Carl Hans, and Klaus-Dieter Thoben, presents approaches for the implementation and configuration of different data integration concepts. Aspects as the speed, comprehensiveness, reliability, flexibility and adaptability of the data integration mechanisms are factors that can limit the degree of possible autonomous cooperation. A systematic analysis of today’s limitations and a categorisation for identifying adequate data integration approaches for different degrees of autonomous cooperation conclude this chapter.
References 1. Scholz-Reiter B, Freitag M, de Beer C, Jagalski T (2006) The influence of production networks’ complexity on the performance of autonomous control methods. In: Teti R (ed) Intelligent computation in manufacturing engineering 5. Proceedings of the 5th CIRP international seminar on computation in manufacturing engineering (CIRP ICME ’06), University of Naples, Naples, Italy, pp 317–320 2. Windt K, Hülsmann M (2007) Changing paradigms in logistics – understanding the shift from conventional control to autonomous cooperation and control. Understanding autonomous cooperation & control – the impact of autonomy on management, information, communication, and material flow. Springer, Berlin, pp 4–16 3. Böse F, Piotrowski J, Windt K (2005) Selbststeuerung in der Automobil-Logistik. Ind Managt 20(4):37–40
194
J. Piotrowski and B. Scholz-Reiter
4. Gehrke JD, Herzog O, Langer H, Malaka R, Porzel R, Warden T (2010) An agent-based approach to autonomous logistic processes. Künstliche Intelligenz 24(2):137–141 5. Böse F, Piotrowski J, Scholz-Reiter B (2009) Autonomously controlled storage management in vehicle logistics – applications of RFID and mobile computing systems. Int J RF Technol Res Appl 1(1):57–76
Chapter 15
Implications of Communication Constraints for the DLRP in Transport Logistics Bernd-Ludwig Wenning, Christian Zabel, Henning Rekersbrink, Carmelita Görg, and Bernd Scholz-Reiter
15.1 Introduction Autonomous control of logistics objects [3] is a new paradigm that changes the way the logistic challenges are handled, away from a centralized planning and dispatching towards considering the logistic objects as autonomous entities that determine their own path through a logistic network. In transportation logistics, this means that the logistic objects such as vehicles and goods become capable of making their own decisions about their route and taking action to follow them. As an autonomous control approach, the Distributed Logistic Routing Protocol (DLRP) has been proposed. Logistic objects that implement this approach use it to collect necessary information about the current status of the logistic network to be able to make sensible decisions. The DLRP specifies the messages and their flows between the logistic objects. Naturally, resources such as bandwidth, energy or memory are not unlimited, and such limitations impose constraints to the communication and interaction of autonomous logistic objects. Communication costs impose another constraint. Therefore, it is necessary to investigate the use of autonomous control concepts such as DLRP under the presence of these constraints to identify the challenges that are encountered. This article focuses on the aspect of how much traffic is generated on the communication interface of an autonomous logistic object that uses DLRP as its autonomous control method.
B.-L. Wenning (B) and C. Görg Communication Networks, University of Bremen, Bremen, Germany e-mail:
[email protected] C. Zabel, H. Rekersbrink, and B. Scholz-Reiter Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany
M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_15,
195
196
B.-L. Wenning et al.
15.2 Related Work: The DLRP As a concept for autonomous control of logistic objects, the Distributed Logistic Routing Protocol (DLRP) [4, 7] has been proposed. It is a decentralized routing method that is based on the assumption that the logistic objects (vehicles and goods) are equipped with devices capable of computing and communicating, or they are represented by software agents acting on behalf of them in a multi-agent system. Thus, there are either physical devices or software agents that enable the logistic objects to interact and decide autonomously. In addition to the vehicles and goods, the vertices (e.g., logistic distribution centers) also become participants in this interaction. In contrast to the classical logistic routing scenarios where heuristic methods are applied to solve static optimization problems such as the Vehicle Routing Problem (VRP) or the Pickup and Delivery Problem (PDP), the DLRP and the scenarios where it is applied are different. First, the DLRP is not a concept to solve static planning problems and create a plan before the transportation process is executed, but it is a control concept that makes real time decisions during transportation process. This way, it is easier to integrate new transport orders into the ongoing process as and when they appear. Further, the scenario topologies are restricted on existing connections between locations (vertices) in the network to better represent existing infrastructure. Scenario topologies are not only defined by a set of vertices, but also by a graph connecting those. The example scenario described later illustrates this. The vertices within such topologies represent, e.g., logistic distribution centers, and the edges represent major road connections between them. Vehicles and goods that use the DLRP determine their routes individually and subsequently match with goods and vehicles that share the same route. The routing is done by using a route discovery messaging that is similar to source routing methods in ad-hoc communication networks: The vehicles/goods send out a route request to the nearest vertex, which forwards it to the neighbor vertices, which in return do the same. Each vertex adds local knowledge about the current network status and transport demand to the request, so that by the time when the request reaches the destination vertex, it has collected information about the complete route that it has traveled. The destination vertex sends a reply to the vehicle or the good, which then can make a decision. After having made a decision, the vehicles and goods announce their intended routes to the involved vertices, where they can be used to create the relevant information for route discoveries from other vehicles and goods. Therefore, the vertices can be considered to act as information brokers. This interaction among vehicles, vertices and goods is depicted in Fig. 15.1. Four main message types are present in DLRP: Route requests (route request messages), being sent to discover routes. Route replies (route reply messages), reply messages returned by the destina-
tions. Route announcements, being sent to publish route decisions. Route disannouncements, being sent to cancel route announcements when a
decision is changed.
15
Implications of Communication Constraints for the DLRP in Transport Logistics
197
Fig. 15.1 Interaction among goods, vertices and vehicles in DLRP [4]
Out of these four message types, the route request is the one that potentially has the largest contribution to the communication traffic created in DLRP as several route requests are forwarded among the vertices in the network in each route discovery. Multi-criteria routing as it is done in DLRP significantly differs from ad-hoc routing protocols in communication networks such as AODV (Ad-hoc On-demand Distance Vector) [2] and DSR (Dynamic Source Routing) [1]: The quality of a route is not necessarily correlated with the sequence of arrival of route request messages at a specific location in the network. The sequence of route request arrivals just allows a statement about the communication path on which the route request has travelled, but not on the logistic route associated with it. Therefore, the common assumption in AODV that the first incoming route request is representing the best route, permitting to drop all the subsequent incoming route requests from the same route discovery, does not hold for logistic routing where the routed goods and vehicles travel in a network whose characteristics are different from those of the communication network that passes the routing messages. Consequently, multiple route requests may need to be processed and forwarded by a vertex during a single route discovery, and multiple route replies may need to be sent back to the logistic object as well. This leads to a potentially high amount of routing messages.
15.3 Possible DLRP Architectures As mentioned before, the DLRP can be implemented in physical devices that are attached to the logistic objects, or in software agents that act on behalf of them. Different levels of distribution, and consequently, different architectures are possible, for example:
198
B.-L. Wenning et al.
Each logistic object is represented by an agent in a central agent system. All
interactions among logistic objects and all decisions are being done within the agent system. The physical goods are only tagged with bar codes or RFID so that they can be tracked whenever they are handled (e.g., loaded or unloaded). The vehicles or their drivers receive the route decisions, e.g., on an onboard unit or handheld device. This setup has a low level of distribution as all decisions are being done on the agent system. Of course, it is also possible to have multiple interconnected agent systems, in which case the level of distribution is slightly higher. The maximum level of distribution would be implemented if each individual logistic object (goods as well as vehicles) is equipped with an individual device that interacts with other devices and makes decisions. Each of these devices acts according to the DLRP specifications. An intermediate level of distribution is implemented if vertices and vehicles run local agent systems. In this setup, the goods are represented by mobile agents that migrate between the agent systems. Decisions are made within the local agent systems on vehicles and vertices. The different architectural options lead to different requirements on the technical side, such as high computational power of the agent system in the central agent system approach, and bandwidth requirements in the approaches with a high level of distribution. In this publication, the focus is on the distributed approach where each logistic object carries its own device.
15.4 Message Sizes Constraints such as costs, bandwidth or memory requirements that may impose limitations on DLRP can be directly related to communication traffic. Therefore, the communication traffic created by DLRP has to be analyzed in terms of message numbers and message sizes. An indication on how much communication traffic is to be anticipated and acceptable is provided by the communication volume that current telemetry systems for trucks generate: Those systems, that usually communicate via GPRS, send data in the order of magnitude of some 100 Kilobyte per month. If DLRP would generate significantly more traffic on the logistic object’s communication link, this would probably not be considered economically reasonable. So the target volume would be in the range of several hundred Kilobyte per month. In the following, the sizes of the four mentioned DLRP message types are discussed. This discussion is based on the assumption that the logistic objects use a decision system as described in [6, 8] with the multiplicative Multi-Criteria Context-based Decision function (MCCD) involving 3 appropriately scaled criteria. The actual sizes of DLRP messages are dependent on implementation details and on how detailed the carried information should be. However, based on the minimal set of information that needs to be included, lower bound estimations are
15
Implications of Communication Constraints for the DLRP in Transport Logistics
199
done in order to investigate traffic volumes that are created by DLRP. The following assumptions are made:
Addresses of logistic entities are 32-bit integers. MCCD is used for the route decisions. 3 context criteria are used in the decisions. Each criterion is described by an identifier, and there are weights and scaling functions for each criterion, as well as limits for acceptable ranges of each criterion. 3 scaling parameters are required per criterion to describe the scaling function that is applied. Because of these assumptions, the calculated message sizes have to be regarded as examples. If more criteria or more complex scaling functions are used, the messages could become larger.
15.4.1 Route Request Messages Route request messages have to contain at least the following information: Vehicle/good flag to identify what type of logistic object is the origin of this
message (boolean, 1 bit). DLRP message type (2 bits). Sender address (integer, 32 bit). Destination address (integer, 32 bit). Sequence number (short integer, 8 bit). Time to live (short integer, 8 bit). Due time (integer timestamp, 32 bit). Good size or vehicle capacity (floating point, 32 bit). Hop list with multiple elements that consist of: ı Hop address (integer, 32 bit). ı Expected arrival time (integer timestamp, 32 bit). ı Expected leave time (integer timestamp, 32 bit). Context field, describing the three criteria and their values, with the following contents: ı Context criteria identifier (3 8 bit ! 24 bit). ı Scaling type identifiers (3 8 bit ! 24 bit). ı 3 scaling parameters per criterion (9x floating point, 32 bit ! 288 bit). ı Context weights (3 floating point; 32 bit ! 96 bit). ı Context limits, one for each criterion and one for the overall route metric (4 floating point; 32 bit ! 128 bit). ı Context values (3 floating point; 32 bit ! 96 bit).
200
B.-L. Wenning et al.
When these numbers are summed up, the resulting message size is 803 C 96n bit, where n is the number of hops. Padding can be applied to create complete octets. Including the padding, the message size is 101C12n Byte. Of course, this is the size at application level, and lower layer protocol (e.g. TCP/IP) overhead is not included.
15.4.2 Route Reply Messages If the destinations just copy the route request into a route reply, the message size remains the same, only the content of the message type field is changed, the time to live field is replaced by a hop count field and source and destination may be swapped. This, however, is not the optimum way as it can be assumed that the vehicle or package remembers with which parameters it has initiated the route discovery, so that the context field does not need to contain all information. If the route reply is cut down to the absolutely necessary information, the following contents remain:
DLRP message type (2 bit). Sender address (integer, 32 bit). Destination address (integer, 32 bit). Sequence number (short integer, 8 bit). Hop count (short integer, 8 bit). Hop list with multiple elements that consist of: ı Hop address (integer, 32 bit). ı Expected arrival time (integer timestamp, 32 bit). ı Expected leave time (integer timestamp, 32 bit). Context field with the following contents: ı Context values (3 floating point; 32 bit ! 96 bit). In this case, the resulting message size is 178 C 96n bit, or (with padding) 23 C 12n Byte.
15.4.3 Route Announcement Messages In the route announcements, the logistic objects inform a vertex about their planned arrival and departure. For this, the announcement has to contain at least the following fields:
Vehicle/good flag (Boolean, 1 bit) DLRP message type (2 bit) Sender address (integer, 32 bit) Announcement number (short integer, 8 bit) Expected arrival time (integer timestamp, 32 bit) Expected leave time (integer timestamp, 32 bit)
15
Implications of Communication Constraints for the DLRP in Transport Logistics
201
Next hop address (integer, 32 bit) Good size or vehicle capacity (floating point, 32 bit) Route preference (floating point, 32 bit)
The sum of these fields results in 203 bit, which corresponds to 26 Byte with padding.
15.4.4 Route Disannouncement Messages A route disannouncement cancels a previous announcement, so it has to at least contain enough data to uniquely identify the announcement that it refers to. Therefore, its contents are:
Vehicle/package flag (Boolean, 1 bit) Message type (2 bit) Sender address (integer, 32 bit) Announcement number (short integer, 8 bit) Next hop address (integer, 32 bit)
This result in 75 bit, leading to 10 Byte with padding.
15.5 Communication Volumes In one route discovery, multiple route requests are propagating through the network. This amount is denoted as NRREQ . However, only one of them is sent by the logistic object that initiates the route discovery, the others are forwardings among the vertices. The destination may send NRREP route replies back to the logistic object. Additionally, route announcements are sent by the logistic object after the route decision (n announcements for a route with n hops), and each route is eventually disannounced with the same amount of disannouncements. When the technical and cost constraints of communication are considered in a fully distributed architecture, the communication traffic has to be categorized in two categories: the first category is vertex-to-vertex communication. This can be considered to be done on wired broadband media. The second category is the communication between vertices and logistic objects. This would usually be done wirelessly, as the logistic object is mobile. Separating the route discovery traffic into these two categories results in the following: Vertex $ Vertex: NRREQ 1 route requests
Vertex $ Logistic object: 1 route request NRREP route replies
202
B.-L. Wenning et al.
n route announcements n route disannouncements
It can be seen that the amount of route requests mainly influences the vertex-tovertex communication, while the other message types are only sent between logistic objects and vertices. Previous work (e.g. [5,8]) has shown that the route discovery in DLRP can create a large amount of communication traffic if no measures are taken to limit this traffic. However, the main focus there has been on the overall network load, where the route requests are the main contribution. Different approaches such as hop count limitation and intermediate route evaluation have been proposed to limit the amount of route requests. Considering the categorization of traffic as shown above, the most constrained link is, however, most likely the logistic object’s wireless link. Per route discovery, this link has to carry, assuming a length of n for all replied routes: 113 Byte .first hop route request/ C NRREP .23 C 12n/ Byte .route replies/ C 26n Byte .route announcements/ C 10n Byte .route disannouncements/ D 113 C 23NRREP C n.36 C 12NRREP/ Byte .total/ It can be seen that the amount of route replies has a significant influence on the communication traffic that the logistic object has to handle. Therefore, an efficiently configured DLRP concept should not only limit the generated amount of route requests, but also the amount of route replies during a route discovery. The simulation results in the following section consider both.
15.6 Simulations Previous work on the traffic volume in DLRP, as described in [5, 8], utilized a scenario topology with 18 vertices based on a German road network. This scenario has now been extended to a topology with 40 vertices and a corresponding set of connections. This topology is still based on a highway network connecting German cities, but with a higher level of detail, as depicted in Fig. 15.2. The vertices labeled as AD and AK are highway intersections which are neither sources nor sinks. All others are sources and sinks. The source and destination locations of transport goods are uniformly distributed among all those vertices, i.e. the mean rate of goods generation is the same in each vertex, and each vertex should also receive the same amount of goods. The overall generation rate is 25 goods per time unit. Each of the goods has to be delivered within 24 time units after its generation. An initial budget is assigned to each of the goods when it is generated. This initial budget is proportional to the shortest distance (in km) between source and destination. Twenty
15
Implications of Communication Constraints for the DLRP in Transport Logistics Bad O.: Bad Oeynhausen Lübeck W.: Wuppertal Heilb.: Heilbronn Hamburg
203
Rostock
AD
Bremen AD
Bielefeld Dortmund Essen Duisburg W. Düsseldorf Cologne Bonn
Berlin
Hanover
Osnabrück Bad O.
Magdeburg Leipzig
Kassel Siegen AD
Erfurt AK
Frankfurt Würzburg
Trier
Dresden
Mannheim
Chemnitz Hof
Nuremberg
Saarbrücken AK Heilb. Karlsruhe Stuttgart 50 km
Ulm
Munich
Fig. 15.2 40-Vertex topology
vehicles, each with a capacity of 12 transport goods, are assumed to be present in the scenario. Each of these vehicles has the same capacity and travel at the speed of 100 km per time unit. In [5], two methods of restricting the amount of goods-initiated route requests were presented and investigated for the 18-vertex topology: forwarding restrictions based on intermediate route evaluation and the application of fixed, scenario-wide hop limits for the route request propagation. These fixed hop limits are now applied to the 40-vertex scenario. Additionally, the route request forwarding is also limited by the route costs. This limit is kept at 50% of the transport good’s current budget. If a transport good does not receive any route reply after a timeout, the route discovery is tried once again with the cost limit dropped, but keeping the hop limit unchanged. Table 15.1 shows simulation results for this setup. Each row in this table shows average values of 5 simulation runs with different random seeds being used for the generation of the goods. In each simulation, 50,000 transport goods were created. The simulation is stopped shortly after the generation of the last transport good. It can be seen that in this scenario, a hop limit of 8 hops is the lower limit. With a limit of 7 hops, less goods are delivered while the amount of route requests increases. The reason is that the network diameter is 8 hops, i.e. a minimum of 8 hops is needed for a connection between the most distant vertices in the topology. A hop limit that is greater than 8 just increases the communication traffic, particularly the amount of
204
B.-L. Wenning et al.
Table 15.1 Route request amounts with fixed, scenario-wide hop limits Hop limit Delivered Goods route Goods route Goods route transport goods discoveries requests replies 7 48,565 1:02 106 8:64 108 2:91 106 8 49,482 6:43 105 4:22 108 4:78 106 5 8 9 49,479 6:43 10 6:33 10 7:19 106 5 8 10 49,480 6:38 10 9:07 10 9:61 106 Table 15.2 Results for adaptive limits Hop limit Delivered Goods route transport goods discoveries 3 C Retry 49,987 1:36 106 counter 4 C Retry 49,986 8:56 105 counter 6 C Retry 49,820 6:77 105 counter
Goods route announcement 2:10 106 2:22 106 2:23 106 2:21 106
Goods route requests 5:94 108
Goods route replies 3:51 106
Goods route announcement 1:78 106
3:51 108
2:29 106
1:85 106
2:96 108
2:44 106
2:13 106
route requests propagating through the network and the amount of route replies sent back to the logistic objects, without much impact on the logistic performance, which is here measured by the amount of delivered goods and the vehicle utilization. So the hop limit of 8 is not only the minimum, but also the optimum if fixed hop limits are applied. The next logical step is to introduce an adaptive hop limit. The adaptive limitation in the given case was chosen as follows: A transport good starts its route discovery with a small hop limit and the cost limit as described above. If the route discovery fails and a timeout occurs, the transport good initiates a retry, increasing the hop limit by one. In case of further failures, this is repeated until the upper limit of the hops is reached. In this setup, the upper limit was chosen to be 9 hops. This method is very similar to the expanding ring search done in AODV. The last retry is done without cost limit to ensure that a route will be found. For the results in the following table, 5 simulation runs were done again for each row, and the amount of created goods was 50,000 again. Table 15.2 leads to the following findings: The trade-off between the number of trials and the communication volume per single discovery can clearly be seen. When the goods are starting with a short hop limit such as 3, they in average need many retries, which is visible from the amount of route discoveries. When they start with a limit of 6 or use a fixed limit of 8 as shown in Table 15.1, there is an overhead caused by increased volume in single route discoveries, some of which are unnecessarily long routes (as visible from the amount of route announcements). With respect to the amount of route replies, an initial hop limit of 4 has turned out to be the optimum, but it leads to slightly more route discoveries as some retries are needed to discover long routes. Starting with a low hop limit leads to slightly more efficient goods routes, so that more goods have been delivered during the simulation time window. The reason
15
Implications of Communication Constraints for the DLRP in Transport Logistics
205
for this is that a transport good makes a decision when it has received a certain amount of route replies without waiting for more. The lower limit eliminates long and inefficient routes, so that it is more likely that the transport good has received a reply with an efficient route within the hop limit. For these reasons, the configuration with an initial hop limit of 4 is considered to be the optimum choice here. In this setup, the amount of messages still seems to be a high number, but this is the cumulative amount for all goods. Per good, this corresponds to 17.1 route discoveries, 45.8 route replies, and 37.0 route announcements (and the same amount of disannouncements). Based on above calculations, each of the route replies has 119 Byte or less as the network diameter is 8 hops. Each first-hop route request has 113 Byte, each route announcement 26 Byte and each route disannouncement 10 Byte. The cumulative communication volume per good in average is therefore: 17:1 113 Byte .first hop route request/ C 45:8 119 Byte .route replies/ C 37 26 Byte .route announcements/ C 37 10 Byte .route disannouncements/ D 8714:5 Byte .total/ This shows that the average communication volume that the individual goods have to handle is well below 10 kilobyte for this scenario. As mentioned earlier, this is of course the volume at application level, so lower layer communication protocols will add some overhead to it. Still, this volume can be considered to be reasonable in an autonomous logistics scenario. However, further evaluations have to be done on larger scenarios to check whether the traffic volume on the logistic object’s link scales well with increasing scenario size.
15.7 Conclusion and Outlook In a fully distributed architecture, DLRP needs to be designed efficiently in order to keep the costs and the resource consumption on an economically and technically reasonable level. This article has investigated the communication traffic that is generated, with a special focus on the logistic object’s communication interface. It has been shown that in the investigated setup and logistic network topology, the communication volume on this interface is in the order of some kilobyte, which is a reasonable volume. On larger topologies, however, it is expected that more communication traffic will be necessary, so further investigations are required here. In future work, this analysis will be done in more detail for different topology sizes to identify whether there are limitations to the size of the logistic network that can be handled reasonably with DLRP. Different architectural options will also be analyzed in detail with respect to the requirements on computational power, communication and memory.
206
B.-L. Wenning et al.
References 1. Johnson D, Maltz D (1996) Dynamic source routing in ad-hoc wireless networks. In: Imielinski T, Korth H (eds) Mobile computing, pp 153–181 2. Perkins C, Belding-Royer E, Das S (2003) Ad-hoc on-demand distance vector (AODV) routing. IETF RFC 3561 3. Scholz-Reiter B, Windt K, Freitag M (2004) Autonomous logistic processes – new demands and first approaches. In: Proceedings of the 37th CIRP international seminar on manufacturing systems, pp 357–362 4. Scholz-Reiter B, Rekersbrink H, Freitag M (2006) Internet routing protocols as an autonomous control approach for transport networks. In: Proceedings of the 5th CIRP international seminar on intelligent computation in manufacturing engineering, pp 341–345 5. Wenning B-L, Rekersbrink H, Görg C (2009) Scalability investigations on communication traffic in distributed routing of autonomous logistic objects. In: 9th International conference on ITS telecommunication, Lille, France, pp 8–12 6. Wenning B-L, Rekersbrink H, Timm-Giel A, Görg C (2009) Weighted multiplicative decision function for distributed routing in transport logistics. In: 2nd International conference for dynamics in logistics (LDIC 2009), pp 117–124 7. Wenning B-L, Rekersbrink H, Timm-Giel A, Görg C, Scholz-Reiter B (2007) Autonomous control by means of distributed routing. In: Hülsmann M, Windt K (eds) Understanding autonomous cooperation and control in logistics – the impact on management, information and communication and material flow. Springer, pp 325–335 8. Wenning B-L (2010) Context-based routing in dynamic networks (dissertation). In: Advanced studies mobile research center Bremen. ViewegCTeubner
Chapter 16
Embedded Intelligent Objects in Food Logistics Technical Limits of Local Decision Making Reiner Jedermann, Javier Palafox-Albarran, Amir Jabarri, and Walter Lang
16.1 Introduction The supervision of food transportation has to be treated as a special case for the application of autonomous control. Firstly, the necessary temperature monitoring produces a huge amount of data that needs to be processed. Furthermore, the data has to be transferred by wireless communication, which typically operates at 866 MHz for passive UHF RFID or 2.4 GHz for wireless sensor networks. Therefore, the high water content is responsible for the high signal attenuation and communication problems in food products. But on the other hand, if the temperature of each transport unit is traced, it is possible to calculate changes in the product quality or losses in the remaining shelf life. If the delivery to retail stores is planned based on the actual shelf life instead of just a fixed production date, the share of products that fall below the quality acceptance threshold can be reduced [17, 28]. The required individual temperature tracing can only be done by sensors inside the goods. First field tests have shown temperature deviations of several degrees Celsius in typical transports. The spatial position of temperature maxima and minima fluctuates for different transports [19]. The field tests indicate that at least 12 or 20 sensors per container are required to estimate a representation of the spatial temperature profile thoroughly. But, despite of the high amount of sensor data, the logistical planning process requires only very few compressed information, for example the remaining shelf life of each transport unit. Several algorithms for data analysis and reduction have been developed and tested by our research cluster CRC 637 “Autonomous Cooperating Logistic Processes – A Paradigm Shift and its Limitations” in the recent years. These include not only the calculation of shelf life as a function of temperature deviation, but also the prediction of the future temperature course, identification of faulty sensors, and spatial interpolation for points that are not outfitted with a physical sensor. R. Jedermann (B), J. Palafox-Albarran, A. Jabarri, and W. Lang Microsystems Center Bremen, University of Bremen, Bremen, Germany e-mail:
[email protected] M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_16,
207
208
R. Jedermann et al.
One of the basic ideas of autonomous control is to shift decision processes from a central unit to distributed platforms [3]. If the data is processed directly at its point of origin, communication can be dramatically reduced. But how does this approach perform under the boundary conditions of our use case in food transportation? Typically, the temperature monitoring of a high number of probe points is done by battery powered wireless sensor networks, which results in the first technical limit of the application of autonomous control. The available processing power is restricted by energy resources. A decision algorithm, which is implemented on an individual sensor node, has to compete with the radio chip and the sensor element for the battery power. Aim of this paper is to evaluate the energy efficiency of decision algorithms on embedded systems and relate it to other factors such as communication and sensor measurement. But, before doing so, the communication range of active wireless sensors will be considered as the second technical limit of the application of autonomous control. The signal attenuation of water-containing products such as fresh fruits was evaluated during the field tests. A set of about 20 sensor nodes was placed in different trucks and containers as described in Sect. 16.2. In order to reduce the required energy for communication, the use of passive RFID tags will also be considered. But, our laboratory tests in Sect. 16.3 showed that passive communication reacts even more sensitively to water-containing products. The effects of a decentralized implementation of decision algorithms on the total energy balance will be summarized in a final section.
16.1.1 Related Work There has been a lot of effort put in by various research groups in topics related to the remote supervision of food transports, but the outcome has not been linked to an overall system so far. Because of the vast amount of literature only single contributions are highlighted in the following overview. James et al. (2006) summarized in a review article several measurements about spatial temperature deviations in trucks and containers and attempts to model the temperature distribution [15]. Biologists have developed models to predict the effect of temperature deviations on the shelf life of numerous types of foods (cf., e.g., [27]). A research group from Athens showed that if deliveries are planned based on the actual shelf life instead of a fixed best before date, losses by decayed food can be reduced by 10% in average [28]. But their approach has not been integrated into an automated supervision system so far. Such a supervision system requires that the origin and the transportation history is known for every product. A traceability system can be implemented based on passive RFID tags [24]. If temperature data should be read out during the transport, wireless sensor networks with a higher transmission range than passive RFID are required. Wireless
16
Embedded Intelligent Objects in Food Logistics Technical Limits
209
sensor networks are an active research field, especially after the TelosB sensor node [8] with the Chipcon CC240 radio, which supports the 802.15.4 communication standard for low-rate wireless personal area networks [12], came in to the market in 2004. Several communication protocols have been developed, which enable forwarding messages over several hops to a base station (cf., e.g., [1, 10]). However, wireless sensor networks can only cover the communication inside the means of transportation; typically GPRS, UMTS or a satellite link is used to cover the external communication. A protocol standard for access to sensor networks over the World Wide Web was suggested by the Open Geospatial Consortium [4]. Current approaches focus on data storage and management [26], but only rarely on the automated evaluation of temperature data. Although there are several methods for sensor data analysis available, they have been seldom applied to sensor networks or food transportation. These include parameter estimation for non-linear system models [11], fault detection and isolation by artificial neuronal networks [34], and spatial interpolation by the so-called Kriging method [30, 31]. The focus of our research is to integrate automated processing of measured transport conditions into intelligent sensors and container systems. The above listed methods for sensor data analysis were adapted to the use-case of food transport supervision and optimized for execution inside wireless sensor networks. There has been only very little research in this area so far: Umer and Tanin (2010) showed how the statistical dependency of sensor measurements can be calculated in a decentralized way by a network of wireless sensor nodes, which is the required information for spatial interpolation by the Kriging method [29]. Furthermore, some commercial data loggers already include a simplified shelf life model [35]. Although the signal attenuation of radio waves by water containing food products turned out as the key problem during our field tests, studies on signal attenuation by food products are hardly found, except for Ruiz-Garcia et al. (2010) for wireless sensor networks and Clarke et al. (2006) for passive RFID [6, 25].
16.2 Technical Limits of Active Communication Inside Packed Foods Active communication means that both the sender and the receiver are supplied by their own power source. But, for battery powered wireless devices as in our use case, the energy, and thereby the transmission power, is limited. The TelosB wireless sensor nodes provide a transmission power of 1 mW at 2.4 GHz. Due to the signal attenuation, the sensor data had to be forwarded over multiple hops to the base station. We distributed between 18 and 30 TelosB sensor nodes during our field tests inside a truck or 40 feet sea container in order to reproduce a typical scenario for monitoring of food transports. Most of the existing protocols for sensor networks solutions focus on the general case with data messages of arbitrary type, size, and direction of transmission. The BananaHop protocol developed by our research group [16] is optimized to forward
210
R. Jedermann et al.
small data packets with temperature and humidity measurements over multiple hops to a base station as required by food transportation supervision. By reducing the scope of operation to this basic task the BananaHop protocol is more energy efficient than the SensorScope protocol [1]. Furthermore, the BananaHop protocol is used as an experimental tool for recording radio signal strength and duration of the active radio period. Although the raw data transfer rate is 250 kbps (Kbit per second) according to the 802.15.4 protocol specification [12] for 2.4 GHz, the effective rate is typically much lower. The transmission of large data packets over a single point-to-point link can theoretically achieve an effective data transmission rate of 101 kbps [21]. If the network can be extended “Ad-hoc” by new sensors at any point of time, timeslots for sending data have to be negotiated anew in each frame. Certain mechanisms for collision avoidance have to be applied. The common CSMA approach [5, page 65] first probes whether the channel is clear. If not, the transfer is delayed for a random period. Control messages to search for routes to the base station require additional channel capacity. Furthermore, the data payload size is rather small in transport supervision scenarios. Only 6 bytes of user data have to be transmitted per frame containing temperature, humidity, and battery voltage measurements. The radio of each sensor has to be powered up for 7.5 s per frame in order to send its own measurement data, wait for an acknowledgment and to forward the data of 30 other sensors by the BananaHop protocol. This example shows that the effective data transfer rate of multi-hop protocols can drop to values as low as 0.2 kbps. The energy consumption for communication was calculated as follows: For sending a single message without acknowledgment the radio has to be powered for 15 ms, requiring 720 J per message at the current consumption of 20 mA at the minimal voltage of 2.4 V for the TelosB. The radio draws almost the same current in receive mode, but because the exact point of time when the message arrives is unknown, the radio has to be powered for an extended period, typically 100 ms, which sums up to a total energy of 5.5 mJ for receive and transmit. The radioup period of the BananaHop protocol in the example above results in an average energy consumption of 360 mJ per frame to transmit one sensor message including overhead for forwarding. The network can operate for 7,000 measurement and communication cycles with a typical battery capacity of 3,000 mAh per sensor. The implementation of the external communication either by GPRS/UMTS or over a satellite link is handled in a separate paper [2].
16.2.1 Experimental Data Losses The application of wireless sensor networks is not only restricted by energy and effective data transfer rate, but most crucially by the communication range, especially if the radio wave propagation is hindered by water-containing products. The radio link quality and the performance of the BananaHop protocol was tested under the conditions of real transports in 2009 and 2010. The field tests were supported by
16
Embedded Intelligent Objects in Food Logistics Technical Limits
211
Dole Fresh Fruit, Cargobull Telematics, and Rungis Express as partners of a transfer project of our research cluster CRC 637. Four separate experiments were carried out in order to analyse the conditions during different parts of the logistic chain, such as short distance truck delivery, long distance sea transportation, and processes inside a warehouse: 1. During the first experiment 20 sensors were installed at the inner walls of a refrigerated truck. Frozen fish and chilled fish products were loaded in two separate compartments. The truck was only partly filled, leaving a free airspace of about 1.5 m above the products. The products were stored inside the truck over the weekend and then delivered to a nearby customer in Bremen, Germany. Only the packet rate of single links was recorded during this first experiment, the full protocol implementation was not yet available. But an analysis of the link data showed that each sensor in each frame had several alternatives to contact the base station either directly or by forwarding over one additional hop. 2. The truck in the second experiment was split in three compartments for different temperature zones. Boxes with frozen meat, chilled fish, and vegetables were stacked at the walls. A corridor in middle of the truck was left empty. The goods were delivered from a distribution centre in the outskirts of Berlin to several customers in the city centre. The whole tour including the return to the distribution centre took 8 h. 3. A third test was carried out inside a banana ripening room at a warehouse close to Hamburg. The sensors were packed in the corners of banana boxes. The boxes with sensors were placed in the centre of the pallets. Ten pallets were loaded into one row of the ripening room. Temperature and link packet rates were recorded over 3 days. 4. The fourth experiment was carried out during 2 weeks of sea transportation of bananas inside a 40 feet refrigerated container from Costa Rica to Hamburg. The densely packed container left only a little free airspace below and above the pallets of 10 cm height. Table 16.1 gives a summary of the four experiments, showing large variations in the percentage of sensor data packets that were not forwarded to the base station (Loss-Rate) and the number of required hops (Max-Hops). The truck tests with a lot of free airspace between the sensors were rather uncritical. At maximum 1.3% of the data messages were lost in one experiment. Three quarters of the sensor nodes could directly send to the base station (1 Hop). The data from the remaining sensors had to be forwarded over one additional sensor (2 Hops). However, if the sensor nodes were packed inside pallets, the data had to be forwarded over up to 5 hops. The experiment in the ripening room showed almost no problems concerning the Loss-Rate. But, inside the packed sea container the LossRate rose to an unacceptable value of 24%. Although the sea container experiment was also the test with the longest duration of two weeks, the length of the experiment cannot be held responsible for the high Loss-Rate. Even if only the first two days of the sea container experiment are considered, the loss rate is with 19% much
212
R. Jedermann et al.
Table 16.1 Summary of experiments with BananaHop protocol Experimental setup
Date
Number of Distance ensors between sensors
1. Truck, partly filled with fish
April 2009
20
1 ... 4m
2. Truck partly filled (mixed load) 3. Banana ripening room 4. Bananacontainer, densely packed
March 2010 July 2009 Sep. 2009
30
1. . . 2 m
18
0.25 . . . 0.5 m
20
0.5 m
Sensor mounting
Max-hops Loss-rate (%)
14 at walls, 6 inside freight At walls
2
0
2
1.3
Inside corner of boxes Inside centre of boxes
5
0.5
5
24
higher as in the truck and banana ripening room tests. The high Loss-Rate is rather caused by the limited free air space inside the loaded container and by the different positions of the sensor nodes inside the boxes. Due to packing, a little free airspace is left in the corners of the boxes, which creates an empty channel in a vertical stack of boxes. The sensors in the ripening room experiment could communicate through this channel, whereas the sensors in the sea container experiment were packed in the centre of the boxes without any free airspace surrounding them. Unfortunately, the last experiment with the highest Loss-Rate is the one that is closest to real-world application. Long distance transports are the most critical for quality degradation, and therefore, the first candidate for sensor supervision. In order to save costs, containers are densely packed. Partly filled trucks are only found in local delivery. Furthermore, the direct core temperature is required for correct quality prediction. Sensors in the corners or close to the surface of the boxes are partly affected by the stream of cooling air resulting in unpredictable behaviour. Tests have shown that small variations in the sensor position can have a large effect on the measured temperature. Some of the sensors mounted close to the surface are mostly affected by the cooling air and cool down with a time constant less than 0.5 days. Others measure rather the fruit temperature and require more than 3 days to cool down.
16.2.2 Analysis of Signal Attenuation in Banana Sea Containers Therefore, the last experiment should be analyzed in more detail. The packet rate was not considered for the network as a whole, but for the direct links between pairs of sensors. The average packet rate of the 12 links with a distance of 0.5 m was 52%. One-third of these links failed completely, another third had temporary dropouts with durations between 8 h and several days, and the remaining third provided almost stable communication. The BananaHop bypassed some of the missing and poor links by an additional hop, but 24% of all messages remained undelivered. Part of these failures is due to inappropriate routing by the network protocol. But
16
Embedded Intelligent Objects in Food Logistics Technical Limits
213
a further analysis of the recorded link information showed that for 20% of all messages there is no physical route available from the source sensor to the base station during the relevant time frame. Therefore, an improvement of the BananaHop protocol or the selection of another protocol could decrease the Loss-Rate by 4% only, but it will not solve the general problem. In order to achieve an acceptable Loss-Rate it is necessary to modify the radio hardware instead. A radio with higher transmission power can be used as the first alternative. The ZigBit Amp OEM Modules from Meshnetics (2008), for example, provide a radio output power of 100 mW, which is 20 dBm higher than that of the TelosB nodes [22]. The other alternative is to use a radio operating at a lower frequency range, which is less sensitive against signal attenuation by water. The signal attenuation is caused by dielectric losses, which are proportional to the imaginary part of the relative electric permittivity ©R . Cole and Cole (1941) provided a formula to calculate ©R as a function of frequency (Equation 1 with ©1 D 6; ©0 D 80; fReso D 16 GHz at room temperature) [7]. "R .f / D "1 C
"0 "1 f 1 C j fReso
(16.1)
For typical frequencies between 433 MHz and 2.4 GHz the imaginary part, and thereby the signal attenuation, is almost proportional to the frequency. Therefore, with regard to signal attenuation, a low frequency would be the best choice. Unfortunately, the selection of the radio carrier frequency is limited to only few ISM (industrial, scientific and medical) radio bands, which are available worldwide or at least in large regions. Large bandwidths are only available at the higher frequencies. For example the ISM band at 2.4 GHz provides a bandwidth of 80 MHz, which can host 16 separate channels for wireless sensor networks with a data rate of 250 kbps each [12]. The ISM band at 915 MHz provides only 26 MHz of bandwidth, but this band is only available in the American continent. For Europe the frequency has to be switched to the ISM band at 868 MHz with only 0.6 MHz bandwidth [12]. The ISM band 433 MHz is not commonly used for wireless sensor networks because the total bandwidth is restricted to 1.6 MHz and the duty cycle is limited to 10% ([12], Annex F). The energy consumption increases for lower bandwidths because the transmission is slower and the radio has to be powered for a longer period of time.
16.3 Limitations of Passive RFID Communication Passive communication needs no power source on the side of the tag and thereby overcomes the energy restrictions of active communication. But on the other hand, the restrictions in communication range and data transfer rate are even more severe. These two limitations of passive RFID were tested in laboratory experiments in order to decide whether passive RFID is a useful alternative to active wireless sensor nodes for the temperature supervision of food products.
214
R. Jedermann et al.
Fig. 16.1 Position of tags at surface of boxes and at bottle necks
RFID tags can be combined with a sensor. Such semi-passive RFID tags contain a battery that is only used to power the sensor and to store the measured values in non-volatile memory. The communication is still passive, meaning that the required energy is provided by the electro-magnetic field of the reader. The reader receives the reflected signal from the tag. Because the radio wave has to cover two times the distance between reader and tag, the signal strength decreases with the fourth power of the distance, not with the second power as in active communication. The experiments were carried out with EPC Generation 2 Tags in the UHF frequency range of 866 MHz. The EPC standard is recommended by the major food retailers. Furthermore, it offers the highest data bandwidth. The signal attenuation by water bottles was tested during the first experiment. Eight tags were placed on the surface of the stack of water bottles and 11 tags behind the first, second, and third rows of bottles (Fig. 16.1, cf. [20]). Reliable identification was possible for the surface tags with a minimum reader power of 200 mW. Tags after the first row required 500 mW of reader power to achieve 100% identification rate. But the tags after the second row achieved only an identification rate of 50%, even at the maximum reader power of 1 Watt. The high sensitivity of passive RFID against water is also supported by other studies. Clarke et al. (2006) observed a reading rate of 97% in a pallet with empty bottles [6]. But, when the bottles were filled with water the reading rate dropped to 0.8%. The reading of RFID tags that are packed inside a container is completely infeasible. Wireless RFID readers can also not be applied for continuous transport supervision. Typically, the RFID reader modules consume between 2 and 30 W. The only practical solution is to read the tags during the loading or unloading of the container. Tags should be mounted to the surface of the product; even 10 cm of water-containing material can reduce the identification rate to an unacceptable value.
16
Embedded Intelligent Objects in Food Logistics Technical Limits
215
If the container is loaded by a fork lift, the data transfer rate becomes a critical factor. The tags are only visible for a short period for the reader antennas. The second experiment was carried out with a pallet of beer bottles placed on a foil wrapper machine. The machine operated at its maximum speed of 10 rotations/min equivalent to an angular velocity of 2.2 km/h. Bulk identification was not problematic; each of the 10 tags at the surface of the pallet could be read at least 29 times/rotation. Figure 16.2 shows the time window for identification. For 1.4 s, more than 90% of the tags could be identified at the maximum reader power. The length of this window, during which the tags are visible to the reader, was compared with the measured data transfer time for different operations in Table 16.2. The time to read 1 Kb of data, equivalent to 700 temperature values with a resolution of 12 bits, could only be estimated based on a projection of the protocol specification, because the semi-passive sensor tags were not available in 2008 when the experiment was carried out.
100 90
1.42 seconds Identifcation rate in %
80 70 60 50 40 30 20 10 0 -2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Time offset in seconds Fig. 16.2 Time window for identification measured on a fail wrapper machine
Table 16.2 RFID data transfer time for different operations Operation Experiment Identification of 4 tags Water bottles, static Reading of 28 data bytes Water bottles, static Reading of 1 Kb temperature data Simulation Writing of 28 data bytes Water bottles, static Writing of 28 data bytes Rotating pallet
Data transfer time (ms) 43 22 172 197 267
216
R. Jedermann et al.
Seven tags each with 1 Kb of memory can be identified and read out during one rotation under optimal conditions. But this rate can hardly be achieved for the reading of sensor tags by a RFID gate during the unloading of a container. Fork lifts typically move faster than the angular speed of only 2.2 km/h as in our experiment. Communication is often distorted and data frames have to be repeated. For this reason we dropped the idea of reading out the full temperature history from semi passive sensor tags. But, the use of low cost RFID technology can become an option if data processing is integrated into the tag. If only the remaining shelf life and the maximum temperature have to be transmitted, a multitude of tags could be read during the unloading and the low data transfer rate would not be an obstacle. This idea of autonomous control on RFID level assumes that tags with programmable micro controller are available, which unfortunately is not the case yet. Therefore, we currently use passive RFID tags only for identification in our project.
16.4 Required Processor Energy for Decision Making Poor radio links can be compensated by repeated transmissions, but this further increases the energy required for communication. The most promising way to reduce the total energy consumption is to minimize the number of data messages by intelligent algorithms that decide which data contains only just redundant information and which contains crucial new information for the logistic planning process. Only the summarized data, the calculated effects of sensor deviations and sensor failure state information, are transmitted instead of the complete measurement data set. But the decision, which data is crucial and which not, needs processing time on the embedded system and thereby energy as well. In order to evaluate the advantages of local decision making, the required energy in (milli) Joule per Decision was evaluated for different example algorithms. However, it is hard to decide whether a local implementation of the algorithm is useful, if only the bare value for energy consumption is known. First of all, the considered algorithms bring clear advantages on their own, independent from their location of the CPU platform, either central or distributed in the network. They perform sensor data evaluation tasks that had to be done manually in the past or have not been done at all. Only with automated processing tools it is possible to carefully analyze the temperature data of 20 sensors from each container. Secondly, it is hardly feasible to directly compare the algorithms because they have different objectives. Two of them predict quality changes or the future temperature development. Two other algorithms for detection of faulty sensors by plausibility checking also have different capabilities: The first focuses on slowly increasing tolerances, and the second one is good in detection of sudden offsets that only affect a single sensor. Finally, the algorithms are executed on different system layers. Some process the data of single sensors; others combine the data of sensor clusters or group of neighbouring sensors, and the rest process the data of the whole container. Some
16
Embedded Intelligent Objects in Food Logistics Technical Limits
Table 16.3 Energy properties of applied wireless sensor nodes Sensor node Processor RAM Flash Typical clock rate Typical supply voltage CPU current (difference between active and idle CPU) Energy per second of calculation
TelosB MSP430 10 Kb 48 Kb 4 MHz 2.4 V 1.5 mA 3.6 mJ
217
iMote2 ARM XScale 32 Mb 32 Mb 416 MHz 3.6 V 50 mA 180 mJ
need only one time initialization at the beginning of the transport and only very few resources to process the data in the following steps. Others are programmed in an incremental form and need almost no initialization. The following section evaluates the advantages of a decentralized local implementation versus a central implementation with regard to energy consumption. Because the required processing power of most algorithms exceeds the capabilities of the TelosB platform, the iMote2 wireless sensor node from Crossbow (2007) was introduced as alternate hardware platform [9]. It uses the same CC2420 radio chip as the TelosB, but provides an ARM XScale processor with much higher computation recourses, but also higher energy consumption. The ARM processor allows the use of more elaborated programming languages as “C#” and “Java” and complex mathematical operations such as large matrix inversions. The TelosB sensors were programmed in “NesC”, which is a special dialect of the “C” programming language. The differences between the two wireless platforms and typical values for supply voltage and clock rate are summarized in Table 16.3. The applied algorithms are briefly introduced before the experimental results of the required CPU times are presented.
16.4.1 Estimation of Temperature Related Effects on Shelf Life In general, the effect of temperature deviations on the quality of foods is of more importance than the temperature itself for the supervision of chilled food transports. The so-called shelf life models [17] were applied to calculate the remaining quality as a function of the temperature history. The considered model applies two Arrhenius type equations to model the temperature dependency of bio-chemical aging and decay processes. The algorithm was programmed in an incremental form, only two exponential functions and two divisions have to be calculated after each temperature measurement. So far, the shelf life model is the only algorithm that has been implemented on the TelosB platform. First tests were carried out in 2008 [17]. Integer operations with 16 or 32 bit were used instead of floating points for faster execution. The average error of integer implementation compared to double precision floating point calculation
218
R. Jedermann et al.
is about 0.5%. The results were compared with a floating point implementation on the iMote2 platform for the current study. Only the resulting remaining shelf life has to be read out at the end of transport instead of transmission of the full temperature history. Alternatively, the system can send a warning message if the shelf life drops below a critical threshold. If the shelf life and the temperature are in range, no communication is necessary at all.
16.4.2 Prediction of Temperature Development The future temperature values inside the container can be calculated by using system identification techniques, which estimate the missing parameters for a given model structure. Online recursive methods require much lower resources in terms of memory and CPU power than offline counterparts, and they are easier to implement on embedded platforms. It is also of paramount importance to have lower order matrix dimensions. The Feedback-Hammerstein parameter adaptation algorithm was implemented on the iMote2 platform [23]. The advantage of the FeedbackHammerstein is that it can also estimate parameters of non-linear effects such as the thermal energy generated by the ripening of the bananas as a function of temperature. Furthermore, it does not need any matrix inversion. In total, three parameters are estimated and updated after each measurement. In order to give an accurate prediction, the model parameters have to be iterated over 3 days at a measurement interval of 1 h, equivalent to 72 cycles. The 3 model parameters were transmitted to the transport operator after this training period. Further communication is only required if the measured temperature deviates from the model prediction. The corrected model parameters and the current temperature have to be retransmitted in this case.
16.4.3 Spatial Interpolation by Kriging Methods to estimate the temperature in points of space as a function of the neighbouring measurements bring further benefits. An estimation of temperature can be necessary for some points because that particular point has no physical sensor at all, the sensor is currently turned off to save energy, or the sensor is unreliable due to faulty measurements. In general, this problem is solved by spatial interpolation. The Kriging method [18] provides an accurate estimation, which is better than simple methods like inverse distance weighting. The measurements of the given sensors (source points) are multiplied with weighting factors to ascertain the temperature in destination points that are not allocated with sensors. The first step of Kriging is to calculate the weighting matrix. It is then applied to the current measurements in the second step. In general, the coefficients of the weighting matrix have to be
16
Embedded Intelligent Objects in Food Logistics Technical Limits
219
calculated only once for each type of transport. The weighting matrix has only to be re-calculated if the loading scheme is modified and the temperature dependency between points change due to changed air streams. A further advantage of the Kriging method is that it can evaluate its own accuracy by calculation of the so-called Kriging Variance as the average error of prediction of the interpolated temperatures [31]. The Kriging method does not work on the level of single sensors as the previous algorithms, but for groups of sensors. In the first test scenario, the Kriging method was implemented on one iMote2 sensor node in order to predict the temperature in 20 destination points by the measurements of 20 source points. Compared to a network that queries all 40 sensors, half of the sensors can be powered down for saving energy. The decision, which sensors should be turned off, can be based on a calculation of the Kriging Variance for their spatial locations. Those with a low Kriging Variance value should have the lowest approximation error for a spatial prediction. But so far, this selection process is not yet automated.
16.4.4 Autonomous Plausibility Checking The risk to draw wrong conclusion due to erroneous measurements increases with the size of the network. Sensors might be faulty because of low battery voltage, mechanical damage, or drifts by aging. Therefore, it is essential to evaluate the reliability of the sensor records. Any abnormality in a wireless sensor network needs to be detected, isolated, and investigated. The measurements of one sensor can be denoted as plausible, partially plausible, partially implausible, or implausible. The deviation can be caused either by a sensor fault or a transport disorder such as unreported opening of the container doors. Plausibility checking needs in general full access to the measurement history. If an autonomous transport monitoring system sends only compressed sensor data such as calculated shelf life and model parameters to the transport operator, the algorithm for plausibility checking also needs to be implemented locally. Two different approaches for plausibility checking in transport supervision were developed by our group. Both compare a prediction for a particular point with the actual measurement at the same location. But, they differ in the way how the prediction is calculated and by their ability to detect different classes of sensor faults. The Kriging method was modified to detect sensors with high tolerances. The actual measurement of each sensor was compared with a spatial interpolation of the remaining 39 sensors in our test scenario. If the residuum between measurement and prediction exceeds the error that should be expected according to the Kriging Variance by a certain factor, a warning message is triggered. The Kriging method detects when a measurement deviates from the typical spatial profile of the temperature distribution. Kriging is well suited to detect high sensor tolerances and slowly increasing offsets. But this approach can only reduce the external communication
220
R. Jedermann et al.
of the container, not the internal one of the sensor network, because the data of all available sensors are required to calculate the prediction.
16.4.5 Artificial Neuronal Networks for Plausibility Checking Artificial neural network (ANN) is a knowledge based technique including nonlinear mapping features and generalization which makes it a favourite for model-free data processing [34]. A plausibility test for clusters of four sensors each was implemented by a multilayer perceptron ANN network [13]. The value for the sensor undergoing the test is predicted based on its own last measurement and the current measurements of its three neighbours. The network consists of two hidden layers with four neurons each and an output layer that sums up the weighted data. Using two hidden layers increases the nonlinear mapping feature between the input pattern and the target. This ANN approach is the best choice to detect sudden changes, for example, by a mechanical damage of the sensor element, battery failure, or intrusion of warm air though an open door. But, slowly increasing offsets cannot be detected because the network simply adapts to it. The weighting factors of the ANN are trained by a modified backpropagation technique. To overcome the memory and processing constraints, the entire network can be updated continuously for training and data approximation solely by using a limited number of neurons and samples, which is called sliding backpropagation [13]. The new algorithm deals with the limitations of wireless sensor nodes in data approximation by using a simplified network. Therefore, the new algorithm is an efficient solution in terms of calculation time and memory size compared to the traditional backpropagation technique. The energy consumption of the sliding backpropagation technique is adjustable by the determination of network architecture, training parameters, and the desired data approximation accuracy. The backpropagation is repeated until the output error drops below a training threshold. The tests were carried out with a medium setting of 0:1ı C. A setting for higher accuracy of 0:001ı C requires three times more CPU resources. The required classification of different fault scenarios is handled by a second ANN. A probabilistic radial basis function (RBF) network can discern between internal sensor faults and external influences [14]. The ANN plausibility testing was implemented under the .Net Micro Framework with C# on the iMote2 sensor nodes and not under Linux with Java as with the other algorithms. The .Net Micro Framework allows only a fixed clock frequency of 104 MHz. The reduced CPU power consumption was considered in the following calculations. Another contribution from our group by Wang et al. (2010) showed that it is also feasible to implement an ANN on the TelosB platform [32]. Sliding backpropagation training for a network with four input neurons can be executed in 162 ms on a TelosB node per step. But, those results could not directly
16
Embedded Intelligent Objects in Food Logistics Technical Limits
221
be compared because its ANN structure differs from the one that was used for plausibility checking.
16.4.6 Dynamic Combination of Algorithms So far, the presented algorithms for autonomous sensor data evaluation were handled separately. But in many application scenarios it would be beneficial to combine two or more of these processes. For example, shelf life prediction and plausibility checking can be combined. Furthermore, if the system detects a faulty sensor by plausibility checking it can apply a third algorithm to replace the missing sensor value by a spatial interpolation. Because it is not known in advance which of the algorithms are required and permanently running, all algorithms create a high overhead; a highly developed system should be able to execute and integrate a number of algorithms only on demand. The system should also provide for the case that the user wants to update an existing algorithm with a new software version or wants to install a new type of data evaluation method. The JAVA based OSGi framework (formerly Open Source Gateway Initiative) was installed on the iMote2 to enable such features [33]. OSGi can update and install the so-called software bundles during runtime without interrupting the execution of the remainder of the system. Except for the ANN based plausibility checking all algorithms are available as OSGi software bundles. Once started, the OSGi framework requires only a little CPU time. Only the installation of new bundles, which takes in average 470 ms of CPU time, has to be taken into consideration.
16.5 Measurement of Required CPU Time The required CPU time for the execution of the described algorithms was measured in laboratory experiments. A digital output pin was programmed to toggle after each model step on the TelosB nodes. The time per incremental model update was measured with an oscilloscope. Timing measurements on the iMote2 nodes were carried out by making use of the system clock. The results are summarized in Table 16.4. CPU times for initialization and incremental update are listed separately. The table also lists the energy consumption for an example scenario where the algorithm is initialized and runs for 50 measurement cycles. For the algorithms, which require spatial data, a setup with 40 sensors distributed on the walls of a delivery truck was considered. Two complete data sets were available from tests in cooperation with a German food provider for hotels and restaurants [18]. The required energy was divided by the number of input sensors for easier comparison.
Only model parameters instead of full temperature history
5 mJ
Only warning for unexpected drop of shelf life
Only warning for unexpected drop of shelf life
Energy for processing 50 sampling intervals including initialization per input sensor Reduced communication
0.17 mJ
Update shelf life 0.96 ms
Incremental operation and CPU time per step
72 steps initial training 360 ms
Update model parameters 5 ms 110 mJ
–
–
iMote2/Java (OSGi) 8.4 Kb 1
Temperature prediction
Update shelf life 0.58 ms
iMote2/Java (OSGi) 7.3 Kb 1
TelosB/NesC (integer) 0.9 Kb 1
Shelf life
Platform/Programming Language Code size Number of input sensors per instance CPU time for initialization
Table 16.4 Comparison of decision algorithms Algorithm Shelf life
279 mJ
Only external communication reduced
About 50% of the sensors can be turned off
Calculation of weighting matrix 58,500 ms Application of matrix 70 ms
Plausibility testing by Kriging iMote2/Java (OSGi) 40.3 Kb 40
28 mJ
Calculation of weighting matrix 2,189 ms Application of matrix 17.5 ms
Spatial interpolation by Kriging iMote2/Java (OSGi) 40.3 Kb 20
Communication restricted to small clusters
Back propagation Training 770 ms 860 mJ
–
iMote2 (104 MHz)/C# 8.4 Kb 4
Plausibility testing by ANN
222 R. Jedermann et al.
16
Embedded Intelligent Objects in Food Logistics Technical Limits
223
16.5.1 Further Energy Consumers The measured values for energy consumption of different algorithms have to be compared with other energy consumers and battery capacity in order to say whether the performance of local decision making is acceptable or not. The sensor measurements were rather uncritical. A typical combined temperature and humidity sensor such as the SHT75 from Sensirion requires only 0.1 mJ per measurement. Only gas sensors cause energy problems because the detector element has to be heated for at least one minute to several hundred degrees Celsius. An example calculation for the AS-MLK sensor for methane from Applied Sensors resulted in an energy consumption of 4,500 mJ per measurement. The stand-by current of the two hardware platforms also has to be considered. The TelosB sensor requires only 1 A when no communication, measurement, or calculation tasks are carried out. But unfortunately, the Linux operating system for the iMote2 does not support low-power deep sleep modes of the ARM processor. The clock frequency can only be switched down to 104 MHz, which still requires a supply current of 55 mA. Before the iMote2 can be applied on transports over several days or weeks, the operating system has to be extended. A low power mode, which halts the CPU but continues to periodically refresh the volatile memory, would require about 1 mA. The energy consumers and the battery capacity are summarized in Table 16.5.
16.5.2 Calculation Versus Communication Energy From an energy point of view, the local decision making algorithms can only be advantageous if the amount of communication is reduced. The reduction of the number of temperature or humidity measurements brings limited benefits only. In the following the total energy for 50 sample intervals was considered in order to attain a fair comparison between algorithms with and without initialization. The algorithms are first compared with a direct send-receive link without the need for
Table 16.5 Summary of energy consumption Operation Sending 50 messages (single direct link transmit and receive/including multi-hop overhead) Decision algorithms (for 50 measurement intervals/one sensor) Measure temperature and humidity 50 times One gas measurement Installation of new OSGi Bundle Standby one day (TelosB/iMote2) Battery capacity (TelosB with 2 AA/iMote2 with 3 AAA)
Required energy 275 mJ/18,000 mJ 0.17 mJ . . . 860 mJ 5 mJ 4,500 mJ 85 mJ 207 mJ/17,100,000 mJ 25,500,000 mJ/11 650,000 mJ
224
R. Jedermann et al. sB
elo
Decision making
l de
fe
li elf
mo
Sh
0.1
1
l de
10
i on
100
Me
1000
Se ure
Energy in mJ
10000
Se
nd
as
Measurement and Communication
te2
Mo
ion ion edict lat r N o p rp e AN mo ty nte ratur i i e l i f i l g e ib elf gin emp us Sh Kri T Pla
T on
nd
/Re
ce
tem
pe
rtu
re
ive
ce
sin
gle
an
ho
p
dh
um
/Re
ive
mu
lti h
op
idit
y
Fig. 16.3 Comparison of energy consumption for 50 intervals
collision avoidance, acknowledgment and other protocol overhead. Such a direct link requires 275 mJ to transmit and receive 50 sensor messages. The resulting energy consumptions are summarized in Fig. 16.3. The shelf life algorithm showed the best performance and the clearest case for local implementation of decision algorithms. The calculation of 50 model steps takes much less energy (0:17 mJ=5 mJ) on both the hardware platforms than sending the same number of sensor messages over a direct link. The shelf life algorithm would break even if the amount of transferred data is cut down by only 1 message, but in fact the communication is almost reduced to zero because only occasional configuration and warning messages have to be sent. The algorithm could run even with the thin-film battery of semi-passive RFID tags [17]. The implementation on the TelosB was 30 times more energy efficient than that on the iMote2. This is partly due to the fact that integer instead of floating point arithmetic was used for TelosB. But, even a floating point implementation on the TelosB requires six times less energy per model step than on the iMote2. The MSP430 processor of the TelosB platform provides the best energy efficiency, but unfortunately it is unable to handle more complex algorithms such as Kriging due to memory restrictions. The local implementation of the model based temperature prediction also gives clear advantages compared to a central solution. After the initial training phase the communication is reduced to occasional parameter updates. The required energy for 50 model steps (110 mJ) is also lower than sending the full temperature data set over a direct link. The Kriging method for spatial interpolation needs much larger amount of energy for the initial calculation of the weighting matrix. But after initialization, the Kriging method replaces the measurements of 20 sensors, not only 1 as in the previous
16
Embedded Intelligent Objects in Food Logistics Technical Limits
225
examples. The related energy consumption per input sensor (28 mJ) is much lower than the saved energy for sensors that can be turned off (275 mJ per sensor). The case for the two plausibility checking methods is a bit more complicated. The plausibility checking by Kriging reduces only the external communication, which is a benefit on its own in regard to the costs of satellite or mobile data transfer tariffs, but this cannot be compared to energy costs. Umer and Tanin (2010) suggested a distributed implementation of Kriging [29]. The weighting matrix does not take the whole network as input, but only 5–10 neighbour sensors. If the matrices are calculated by dedicated sensor nodes inside the network, the internal communication of the sensor network could be reduced as well. The ANN based plausibility checking needs the highest computation resources (860 mJ) per sensor, but it also includes a method for classification of different fault types. Other as the Kriging method, the ANN based approach is well suited for an implementation within the network. Only four sensors of a cluster instead of all 40 have to send their data to a cluster head, which carries out the plausibility checking. The communication of the wireless sensor network can be reduced by 90% in the test scenario. Under the optimistic assumption of a network that consists only of direct links, a distributed implementation of the ANN brings no energy advantages. But the protocol overhead cannot be neglected in a real wireless sensor network. If the energy consumption of the BananaHop protocol is taken as reference (18,000 mJ), the calculation of the sliding backpropagation requires only 5% of the saved communication energy.
16.6 Summary and Conclusions The spatial supervision of temperature and other sensor parameters of transports of perishable goods create a high data volume that is difficult to handle manually. Several automated algorithms for sensor data processing were introduced, which not only summarize the sensor data but also provide additional types of information such as the calculated remaining shelf life or indication of faulty sensors. Because the two suggested methods for plausibility checking have different focuses, it is recommended to implement both of them in order to detect slow increasing tolerances as well as sudden offsets by malfunction of the sensors or external influences. The methods for automated sensor data processing can be implemented either centrally on a server in the office of the transport operator or locally, directly on the sensor nodes or a processing platform inside the means of transportation. But before the local implementation can come into practice, two technical problems have to be solved: Firstly, the high signal attenuation by water-containing products has to be compensated by adapted radio hardware; secondly, the operating system for the iMote2 has to be extended to support low-power sleep modes in order to reduce the stand-by current.
226
R. Jedermann et al.
The measurements of the required CPU times showed that it is feasible to run the algorithms on low-power embedded systems. Decisions can be made at the hardware platform where the input data has its spatial origin. This is especially the case for the temperature supervision of perishable products with a high volume of distributed sensor data. On the other hand, the transport planning of “dry” goods without sensor supervision mainly requires data, which do not origin from the truck or container such as new orders and traffic information. The distributed implementation of sensor data processing mainly brings advantages for the case of perishable goods. The communication volume is reduced, thereby the energy is saved and the system becomes less dependent from communication failures. The methods for shelf life calculation, temperature prediction, and Kriging interpolation save more energy than they require for computation, even under the optimistic assumption that the sensor network can run without any protocol overhead. The case for more complex algorithms for plausibility checking depends on the external communication costs and the internal protocol type. But an energy comparison with typical multi-hop protocols also shows clear benefits. The suggested algorithms can run on the CPU of typical sensor nodes without hardware extensions. If the installation of a wireless sensor network for spatial transport supervision is planned, it is recommended to extend the sensor node software for autonomous data processing concurrently.
References 1. Barrenetxea G, Ingelrest F, Schaefer G, Vetterli M (2008) Wireless sensor networks for environmental monitoring: the sensorscope experience. Paper presented at the 20th IEEE international Zurich seminar on communications (IZS 2008), Zurich, Switzerland, pp 98–101 2. Becker M, Wenning B-L, Görg C, Jedermann R, Timm-Giel A (2010) Logistic applications with wireless sensor networks. In: HotEmNets 2010, Ireland 3. Böse F, Windt K (2007) Catalogue of criteria for autonomous control in logistics. In: Hülsmann M, Windt K (eds) Understanding autonomous cooperation and control – the impact of autonomy on management, information, communication, and material flow. Springer, Berlin, pp 57–72. doi:10.1007/978–3–540–47450–0_5 4. Botts M, Percivall G, Reed C, Davidson J (2007) OGC sensor web enablement: overview and high level architecture. Open geospatial consortium white paper. Retrieved from: portal.opengeospatial.org/files/?artifact_id=25562 5. Callaway EH (2004) Wireless sensor networks – architectures and protocols. Auerbach Publication, London, New York 6. Clarke RH, Twede D, Tazelaar JR, Boyer KK (2006) Radio frequency identification (RFID) performance: the effect of tag orientation and package contents. Packag Technol Sci 19(1): 45–54. doi:10.1002/pts.714 7. Cole KS, Cole RH (1941) Dispersion and adsorption in dielectrics. J Chem Phys 9:341–351 8. Crossbow (2005) TelosB mote platform. Product datasheet. Retrieved from: http://www. willow.co.uk/TelosB_Datasheet.pdf 9. Crossbow (2007) IMote2 – high-performance wireless sensor network node. Product datasheet. Retrieved from: http://www.xbow.com/Products/Product_pdf_files/Wireless_pdf/ Imote2_Datasheet.pdf
16
Embedded Intelligent Objects in Food Logistics Technical Limits
227
10. Demirkol I, Ersoy C, Alagoz F (2006) MAC protocols for wireless sensor networks: a survey. IEEE Commun Mag 44(4):115–121 11. Guo F (2004) A new identification method for Wiener and Hammerstein systems. FZKA 6955. Forschungszentrum Karlsruhe, Germany 12. IEEE Computer Society (2006) Wireless Medium Access Control (MAC) and Physical Layer (PHY) Specifications for Low-Rate Wireless Personal Area Networks (WPANs). vol Standard 802 Part 15.4 13. Jabbari A, Jedermann R, Muthuraman R, Lang W (2009) Application of neurocomputing for data approximation and classification in wireless sensor networks. Sensor J 9:3056–3077 14. Jabbari A, Kreowski H-J, Lang W (2010) Application of bio-inspired data processing in intelligent transportation systems. Paper presented at the intelligent systems (IS), 2010 5th IEEE international conference, London, UK, 7–9 July 2010, pp 315–320. doi:10.1109/IS.2010.5548366 15. James SJ, James C, Evans JA (2006) Modelling of food transportation systems – a review. Int J Refrigeration 29(6):947–957. doi:10.1016/j.ijrefrig.2006.03.017 16. Jedermann R, Becker M, Görg C, Lang W (2011) Testing network protocols and signal attenuation in packed food transports. Int J Sensor Netw (IJSNet) 17. Jedermann R, Edmond JP, Lang W (2008) Shelf life prediction by intelligent RFID. Paper presented at the dynamics in logistics. first international conference, LDIC 2007, Berlin/Heidelberg, pp 231–238. doi:10.1007/978–3–540–76862–3_22 18. Jedermann R, Lang W (2009) The minimum number of sensors - Interpolation of spatial temperature profiles. Paper presented at the wireless sensor networks, 6th European conference, EWSN 2009, lecture notes in computer science (LNCS), Berlin/Heidelberg, pp 232–246. doi:10.1007/978–3–642–00224–3_15 19. Jedermann R, Moehrke A, Lang W (2010) Supervision of banana transport by the intelligent container. Paper presented at the Coolchain-management, 4th international workshop, Bonn, Germany, pp 75–84 20. Jedermann R, Stein K, Becker M, Lang W (2008) UHF-RFID in the food chain – from identification to smart labels. In: Kreyenschmidt J (ed) Coldchain manangement. 3rd International Workshop, Bonn, pp 3–15 21. Jennic (2006) Calculating 802.15.4 data rates. Application Note JN-AN-1035 Retrieved from: http://www.jennic.com/download_file.php?supportFile=JN-AN-1035+Calculating+80215 4+Data+Rates-1v0.pdf 22. Meshnetics (2008) ZigBit Amp OEM Modules – Ultra-Compact 2.4 GHz 802.15.4/ZigBee Modules with Power Amplifier for Wireless Networking Applications. Product datasheet, Revision 2.2. Available from: http://www.meshnetics.com/zigbee-modules/amp/ 23. Palafox-Albarrán J, Jederman R, Lang W (2010) Prediction of temperature inside a refrigerated container in the presence of perishable goods. In: 7th International conference on informatics in control, automation and robotics (ICINCO), Portugal 24. Regattieri A, Gamberi M, Manzini R (2007) Traceability of food products: general framework and experimental evidence. J Food Eng 81(2):347–356. doi:10.1016/j.jfoodeng.2006.10.032 25. Ruiz-Garcia L, Barreiro P, Robla JI, Lunadei L (2010) Testing ZigBee motes for monitoring refrigerated vegetable transportation under real conditions. Sensors 10(5):4968–4982. doi:10.3390/s100504968 26. Schneider M, Kroner A (2008) The smart pizza packing – an application of object memories. In: 4th International conference on intelligent environments, 2008 IET. Seattle, WA, USA, pp 1–8 27. Tijskens LMM (2004) Discovering the future: modelling quality matters. Ph.D. thesis, University of Wageningen, Netherlands 28. Tsironi TE, Gogou P, Taoukis PS (2008) Chill chain management and shelf life optimization of MAP seabream fillets: a TTI based alternative to FIFO. Paper presented at the Coldchain management. 3rd International workshop, Bonn, Germany,pp 83–89 29. Umer ML, Tanin E (2010) Spatial interpolation in wireless sensor networks: localized algorithms for variogram modeling and Kriging. GeoInformatica 14(1):101–134. doi:10.1007/s10707–009–0078–3
228
R. Jedermann et al.
30. Wackernagel H (2003) Multivariate geostatistics - an introduction with applications. Springer, Berlin 31. Walkowski AC (2008) Model based optimization of mobile Geosensor networks. In: Bernard L, Friis-Christensen A, Pundt H (eds) The European Information Society – taking geoinformation science one step further. Lecture notes in geoinformation and cartography. Springer, Berlin Heidelberg, pp 51–66. doi:10.1007/978–3–540–78946–8_3 32. Wang X, Jabbari A, Laur R, Lang W (2010) Dynamic control of data measurement intervals in a networked sensing system using neurocomputing. Paper presented at the international conference on networked sensing systems (INSS 2010), Kassel, Germany 33. Wessels A, Jederman R, Lang W (2010) Embedded context aware objects for the transport supervision of perishable goods. Paper presented at the recent advances in elektronics, hardware, wireless and optical communications, Cambridge, pp 119–125 34. Zaknich A (2005) Principles of adaptive filters and self-learning systems. Advanced textbooks in control and signal processing. Springer, Berlin 35. Zweig SE (2008) LifeTrack technology for smart active-label visual and RFID product lifetime monitoring. Paper presented at the coldchain management, 3rd international workshop, Bonn, Germany, 29–36
Chapter 17
Knowledge Management for Agent-Based Control Under Temporal Bounds Tobias Warden, Robert Porzel, Jan D. Gehrke, Hagen Langer, Otthein Herzog, and Rainer Malaka
17.1 Introduction In recent years, multi-agent systems have established themselves as a tried and tested approach [13] that lends itself particularly well for applications in domains that exhibit the following characteristics. One such characteristic is that an opportunity exists to decompose a given complex problem into a number of less complex sub problems. Another characteristic is that a natural mapping can be found from real-world actors to software agents that are made responsible for finding solutions for the individual sub problems, such that the agent community as a whole generates a solution for the initial complex problem. An additional case where multi-agent systems can be considered a suitable means for representation, modeling and simulation of a given state of affairs is found when multiple stakeholders that each seek to achieve their respective logistic objectives compete or cooperate to some degree. Autonomous logistics is an application domain where the characteristics stated above hold. First, the task of reaching a specified logistic goal is naturally composed of individual logistic tasks carried out by multiple real world entities. Second, for each logistic entity there is a more or less straightforward mapping to an individual software agent that acts as a digital representative for that entity. As explored in simulations of transport and supply-chain scenarios (cf. e.g. [21]) the realization of autonomous logistic control systems on the basis of multi-agent systems works well for achieving the overall specified logistic goals using situation-aware software agents that apply their knowledge to fulfill their individual subtasks. However, several challenges remain to be tackled before one can deploy such a multi-agent system outside of the simulation sandbox. That is not to say that these challenges could not arise inside the simulation sandbox, as well. Some of these challenges are tightly intertwined with temporal issues. In applying multiagent systems in autonomous logistics, the temporal ramifications of an agent’s decision have been considered as one of the key logistic factors, e.g., how long a T. Warden (B), R. Porzel, J.D. Gehrke, H. Langer, O. Herzog, and R. Malaka TZI – Center for Computing Technologies, University of Bremen, Bremen, Germany e-mail:
[email protected] M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_17,
229
230
T. Warden et al.
specific delivery takes depending on the chosen route [6]. When enabling individual agents to adopt their domain knowledge and learn throughout their life cycle, i.e., to acquire knowledge, build corresponding models and apply those in their decision making processes, this will have temporal ramifications that need to be considered together with their physical counterparts. In this work, we will spell out possible temporal interdependencies between the knowledge management processes an agent performs and its primary logistic operations. We will thereby focus on the implications these temporal interdependencies have for the different learning and knowledge management operations that an agent can perform.
17.2 The Need for Knowledge Management A natural approach to bring about the paradigm shift from traditional, centralized control of logistic processes to autonomous logistics is the realization of control systems on the basis of multi-agent systems (MAS). In such a system, the constituent software agents act as digital representatives on behalf of logistics entities, which can be broadly categorized in business resources such as means of transport or storage facilities and commodities [10].
17.2.1 Forms of Knowledge Logistic agents need knowledge of different kind in order to handle their respective roles in the modeled logistic processes effectively and in the desired quality of service. The spectrum reaches from declarative knowledge, e.g., ontologically modeled background knowledge about the relevant excerpt of the logistics domain [22] down to operative knowledge for support of agent decision processes [4]. A subset of the knowledge required for the operation of a multi-agent system can be prepared at design time, for instance by means of knowledge engineering where expertise from human agents is acquired, preprocessed and translated to machineunderstandable formats such as formal ontologies. Default models for decision support may be either devised by hand or preferably learned offline from historical data acquired from external sources or gathered specifically in the targeted application domain). These pieces of knowledge can then be rendered available to all actors in a multi-agent system as joint a-priori knowledge. In order to accommodate the complexity and dynamics of logistic environments, the initial provision of default knowledge alone is not sufficient. It is rather necessary to design the employed agents as adaptive systems, capable of autonomous knowledge revision and the compilation of additional empirical or a-posteriori knowledge, such that the available knowledge becomes to an increasing degree tailored to the respective deployment context. To that end, local machine learning
17
Knowledge Management for Agent-Based Control Under Temporal Bounds
231
enables this individualization and the build-up of situated knowledge. Of particular importance in this process is such knowledge which, in combination with corresponding inference mechanisms, immediately supports the primary logistic operation, such that a qualitative improvement of decisions is brought about, for instance, predictive models for traffic densities or trans-loading durations [4].
17.2.2 Agent-Oriented Knowledge Management Agent-oriented Knowledge Management (AKM) refers to a collection of processes that an intelligent agent carries out throughout its life cycle in order to maintain, adapt, and possibly augment its empirical and, to a lesser extent, a-priori knowledge through knowledge acquisition. Moreover, the technical term also comprises processes such as goal-directed information acquisition and the ascertainment of new knowledge by means of inference on either predictive models or decision models. Agent-oriented knowledge management bears analogies to Personal Knowledge Management. While the latter focuses on human, rather than artificial decision makers, both approaches have similarities in the considered skill sets. Examples comprise self-reflection, management of individual learning, information literacy, as well as communication and collaboration. We understand AKM as a form of distributed knowledge management on the part of the individual actors within a multi-agent system [14]. The agents within the multi-agent system may also form knowledge networks and pursue strategies for demand-oriented knowledge transfer. However, in the following we will restrict the focus to knowledge management functions at the level of single agents.
17.2.2.1 Knowledge Management Activities The intelligent agents that we consider contribute to the autonomous control of logistic processes. For the most part, they thereby act as digital representatives on behalf of logistic entities, be they resources of a logistic service provider such as a forwarding agency or the subjects of the considered logistics processes, e.g., pallets or containers in transport or work pieces in production. It is important to note that these agents therefore need to attend to their respective primary logistic activities, right from the moment of their instantiation. Therefore, their knowledge management functions assume auxiliary activities that are loosely coupled to the primary logistic ones. This initial situation bears implications for the organization of agentoriented knowledge management. The agents act under economic considerations. As a consequence, they are constantly faced with the challenging task to reconcile short-run logistic interests, such as achievement of customer satisfaction by delivering pre-negotiated quality of logistic services, and strategic interests, for instance to achieve a competitive advantage over competitors. Broken down to knowledge management, the former interest demands an efficient exploitation of available
232
T. Warden et al.
knowledge and supporting models in decision processes. On the other hand, in order to improve their basis for decision support in the momentary deployment context and retain the agility to quickly adapt to context chances, the agents also require a wealth of diverse individual experience. Metaphorically speaking, indirect paths broaden one’s knowledge of a place. This leads to research questions such as: What are feasible options to incorporate exploration into the operative action of an agent without giving rise to a lowering of logistic service quality? Besides such questions, the respective logistic circumstances also constrain the periods of time agents are granted for the completion of particular knowledge management processes. For instance, what would be considered acceptable upper bounds for computation times with regard to the construction of a new decision support model or inference on an existing model in a concrete decision situation? In addition, it is necessary to reduce the context-dependent choice of when and in which frequency model revisions ought to be prompted. Before broaching these research questions further, we will characterize in more detail the knowledge management functions sketched out above. To begin, we can assess that these functions are executed due to independent incentives.
17.2.2.2 Knowledge Management in Concrete Decision Situations First, in the face of a concrete decision situation, agent-oriented knowledge management enables the agent as autonomous decision maker to gather, and process raw data from approachable sensors, deployed for instance on its managed physical object as in the case of an intelligent container [11]. By means of interpretation in context, this raw data can then be elevated to information. Conjointly with supplemental information acquired from cooperating agents or other external information sources, such as EPC1 Information Services (EPCIS) or enterprise resource planning (ERP) systems [10], the gathered information is filtered for relevance and the remainder integrated with the belief base of the respective agent. Subsequently, the knowledge can then be exploited. This may happen immediately in deliberation. In our application domain of transport logistics, such processes involve, for instance, the planning of multi-tier haulage of commodities within supply networks [21] or the planning of a forwarding agency’s participation in operational transport collaboration [2]. Alternatively, the gathered information may provide necessary input for the agent’s decision support models (for instance as evidence in Bayesian networks used as predictors). These are typically queried in the decision process in order to deduce new knowledge by means of inference. In fact, it is often the case that the demand for information acquisition arises as a direct consequence of the desire to perform inference on decision support models. The knowledge management functions described above, therefore, have as a common denominator that they are triggered by the agent’s primary logistic activity.
1
Electronic Product Code.
17
Knowledge Management for Agent-Based Control Under Temporal Bounds
233
17.2.2.3 Knowledge Management with Strategic Background A second set of knowledge management functions revolves around the initial construction of tailored decision support models at the outset of an agent’s life cycle and their subsequent adaptation in the light of new experiences. These functions contribute to a sustainable knowledge management. They are therefore typically performed detached from particular decision situations the agents may encounter. As a consequence, it is often feasible to perform these functions concurrent to the execution of the agent’s primary logistic roles. The absence of trigger events for this second category of knowledge management functions out of the operational day-to-day business implies that their scheduling needs to be managed pro-actively by the agent. Thereby, different strategies and prioritization are conceivable. Their selection is, as we will develop in throughout this contribution, strongly biased by the agent application domain. In an even more direct way, it is also the case that decision-induced knowledge management functions are, in practice, constrained in design by the agent deployment context.
17.2.2.4 Knowledge Management Under Real-World Constraints In the following, we will therefore develop an integrated perspective on agentoriented knowledge management. We argue that, based on a careful classification of knowledge management processes into the agent life cycle, we enable the derivation of design cues for aspects of agent-based knowledge management. These will particularly account for temporal constraints, such as the limited time that is at the disposal of an agent for learning, adaption, and deliberation. To that end, we work out interdependencies of primary agent role and auxiliary knowledge management functions over the full course of the agent life cycle. Also, we argue that the task adequacy of employed approaches and algorithms in both knowledge acquisition (which comprises, for the purposes of this report, gathering of experience data, data pre-processing and machine learning) and knowledge exploitation (i.e., inference used in particular decision situations) needs to be assessed on an individual basis with regard to temporal constraints given in the respective case.
17.3 Agent Activities and Temporal Constraints In the following we will describe the individual activities of an intelligent logistic agent in the light of their temporal ramifications. In order to find some common ground we first propose a minimal nomenclature and subsequent modeling approach for discussing these activities under temporal constraints.
234
T. Warden et al.
17.3.1 A Minimal Model for Agent Activities In this section we initially explicate a formal framework to investigate automated planning and scheduling for intelligent adaptive agents in dynamic domains such as autonomous logistics. This framework incorporates both activities and tasks that belong to the primary, in our case logistic, agent roles and those which belong to agent-oriented knowledge management. For the purpose of introducing the formal framework, we consider a paradigmatic minimal scenario that makes several simplifying assumptions. The scenario is that of a transport agent that organizes the operative handling of transport orders on behalf of a particular means of transport within a freight forwarder fleet. This agent continually receives new transport orders directly from customers or superordinate agents. Each order thereby corresponds to a specific pickup and delivery activity that needs to be first planned and scheduled, and consequently executed on the operative level. We characterize this activity as atomic activity .ActAtom / under the assumption, that it cannot be decomposed in constituent activities which each require a dedicated planning phase followed by execution, i.e., planning and execution is not, or does not need to be, interleaved. The atomic activity however features an internal structure. In this context, we speak of constituent tasks. Following the account above these tasks comprise: Deliberation tasks that comprise the actual automated planning and scheduling
as particular form of decision-making as well as knowledge management related tasks such as knowledge acquisition and, possibly, the formation of decision support models. Let the universe of these sub tasks be denoted as TaskDel : Execution tasks. In our application domain of logistics, such tasks may encompass transport (including routing), cargo handling, storage and commissioning. We denote the universe of such tasks by TaskExec . The former task can thereby be further broken down into three sub tasks: Data, information and knowledge acquisition .TaskDA / Model formation .TaskDG / Model employment for decision-making .TaskDE /
where model formation, as we shall show, is optional and depends on the learning strategy adopted by the agent. The identified sub tasks may be broken down further as indicated in Fig. 17.1 on the following page. Information, for instance may have to be acquired from different sources such as peer agents or third-party information services. An atomic activity ActAtom as sketched above, is also characterized by two types of temporal constraints, namely (1) ordering constraints and (2) duration constraints. We initially introduce ordering constraints that specify the chronology of a number of tasks. However, they leave unspecified at which point in time the tasks start and terminate and thus how long their respective time intervals are.
17
Knowledge Management for Agent-Based Control Under Temporal Bounds ActAtom
235
tnow
t
and
TaskDel
IDue TaskExec
(m) and
... TaskDA
...
TaskDE
TaskDG
(m)
t
(m)
...
I(ActAtom )
(m,o,f)
I(TaskDel )
(m)
(m)
(m)
I(TaskExec) //
...
...
...
...
...
...
...
...
...
I(TaskDA )
I(TaskDG )
I(TaskDE )
time
Fig. 17.1 Left: Schematic decomposition of an atomic activity ActAtom into constituent tasks. Right: Schematics of associated temporal intervals and their temporal relations (Notation follows [1])
To begin with, pickup and delivery orders are typically due within a time window IDue D ht’ ; t¨ i. Let now I.TaskDel / D ht0 ; t• i denote the time interval in which all deliberation takes place. Let further I.TaskExec / D ht ; t® i denote the time interval in which the execution takes place: The considered agent needs to perform TaskDel before TaskExec . Using notational
means introduced by Allen (1984) [1], the respective ordering constraint can be specified precisely as: Meets .I.TaskDel /; I.TaskExec // _ Before.I.TaskDel /; I.TaskExec //. The agent needs to conclude TaskExec before t¨ since only then is ActAt om
accomplished on schedule: Meets .I.TaskExec /; IDue / _ Overlaps.I.TaskExec /; IDue / _ Finishes.I.TaskExec /; IDue /. The schema on the right of Fig. 17.1 depicts the chronology of sub tasks together with the ordering constraints that apply in the example. If the agent is enabled to form new decision support models inline, i.e., as part of the deliberation process plan for the execution of its current logistic activity, the full fine-grained distinction of TaskDel presented above applies which states that TaskDel is constituted by three knowledge related tasks TaskDA ; TaskDG , and TaskDE . These tasks are also structured by ordering constraints. A simple sequential case could be formalized as: Meets .I.TaskDA /; I.TaskDG // ^ Meets .I.TaskDG /; I.T askDE //. The resulting activity decomposition and the identified temporal ordering constraints can then be visualized by means of an adapted and-or-tree as shown on the left side of Fig. 17.1. The tree representation allows for the classification of learning approaches in terms of their associated tree configuration. In the base scenario that has been introduced as a reference, the model formation conducted by the agent specifically in
236
T. Warden et al.
the context of a particular atomic logistic activity and herein the deliberation task. Now consider the distinction between eager and lazy learning approaches, which is common in the field of machine learning. Eager learning algorithms work by generating hypotheses for all target predicates in advance, i.e., they form an explicit (prediction) model from the provided data. Lazy learning algorithms, by contrast, postpone that hypothesis generation until presented with a concrete query. The distinction between eager and lazy learning is thus related to the ability of lazy learners to generate local approximations of the target hypothesis with respect to the concrete query based on the raw data alone, whereas eager learners have to commit to a single hypothesis at training time [16]. This translates to our model such that in the case of eager learning, the experience that has been collected hitherto through perception and in prior information acquisition .TaskDA / is employed directly in inference during the decision-making .TaskDE /. In eager learning, by contrast, new models are generated beforehand in the TaskDG phase based on experience gathered in preceding TaskDA phases. The resulting tree configurations for both families of learning approaches are depicted in Fig. 17.2. Having introduced ordering constraints in our description framework, the natural next step is the consideration of duration constraints. Specifically, automated planning and scheduling of sub tasks in agent activities needs to incorporate temporal presets from the application domain (e.g., in our basic case, the specification of pickup and delivery intervals). These then lead to constraints with regard to the maximum temporal extension of each node in the trees shown in Fig. 17.2. We argue that with regard to planning and scheduling of agent activities, both primary logistic task and those tasks related to agent-oriented knowledge management should be factored on equal grounds.
ActAtom
ActAtom
and
TaskDel
and
TaskExec
(m)
TaskDel
and
and
...
TaskDA
...
TaskSel
... (m)
...
...
...
TaskDE
TaskDG
(m)
...
TaskDA
(m)
TaskEval
...
(m)
... (m)
...
TaskDE
and
and
(m)
TaskExec
(m)
and
...
...
...
...
...
...
TaskInf
TaskInf
TaskPre TaskLearn ...
...
...
Fig. 17.2 Left: Schematic activity decomposition for an agent employing an eager learning approach to create a model on the fly. Right: Alternative decomposition for an agent that resorts instead to a lazy learning approach. Computational complexity is in this case shifted to the inference task, as no model besides the base data itself is required
17
Knowledge Management for Agent-Based Control Under Temporal Bounds
237
In particular cases, that is when the considered activity has a rather short temporal extension, the situation may arise that decision-making needs to be concluded within a time frame that prohibits computationally expensive comprehensive model formation in order to obtain prediction results. Such cases are conceivable, for instance, in the context of collaborative route planning [2].2 In such a situation, the extrinsic parameters of the logistic tasks thus foreclose time-consuming eager learning approaches as sketched on the left of Fig. 17.2. If, however, the agent is equipped with several learning strategies, it may still be possible to opt for eager learning as in the right-hand side of that same figure. The following Sects. 17.3.2 and 17.3.3 will consider further aspects of agentoriented knowledge management with particular focus on temporal constraints. We thereby split knowledge management processes in two complementary aspects: The first considers strategic knowledge management processes. These are meant to continuously augment the repertoire of decision- and prediction models of an agent. The second aspect is then concerned with model exploitation in concrete decision situations. Knowledge management here immediately enables informed decisionmaking by the agent. This is the underlying use case that we have assumed thus for in the introduction of our formal description framework. Here, knowledge management functions become triggered and are incorporated in deliberation, in particular decision-making, in a concrete decision situation.
17.3.2 Compilation and Maintenance of Models Strategic knowledge management processes comprise the generation and maintenance of knowledge models that support the primary agent activities. The agent typically executes these processes pro-actively in the background, i.e., decoupled from its primary tasks. This is reflected for instance in the abstract architecture of a learning agent [20]. As Fig. 17.3 illustrates, an adaptive agent can be conceptually distinguished into a performance element, which in our use cases comprises the primary logistic activities, and an additional learning element, which hosts the agents facilities to learn and adapt. Although these agent aspects are distinct, they nevertheless interact. On the one hand, all learning by the agent is based on experience acquired through continuous interaction with the logistic task environment. On the other hand, the learning element contributes, via eager learning, the decision support models that enable, in the first place, competitive operation by the performance element. Besides these central architectural components of the learning agents, Fig. 17.3 also introduces two additional components that have high relevance for adaptive logistic agents; namely, the critic and the problem generator. 2 In transport logistics, where the haulage of commodities may take hours, days, or even weeks, the time spans required for decision-making, including planning and scheduling as well as model formation, are by contrast often negligible.
238
T. Warden et al.
Performance Standard Sensors
Critic
1
feedback
Learning Element
2
changes knowledge
Performance Element
learning goals 3 Problem Generator Learning Agent
Actuators
Environment
Fig. 17.3 Abstract architecture of a learning agent (adapted from [20])
The critic is the component that can assess the performance of the agent or, more specific, particular decision (support) models, with respect to an invariant performance standard. In the case of a prediction model, such a standard could be an aspired precision and recall. For decision models the performance standard could be directly derived from logistic service specifications. The situation assessment of the critic, which is based on the agent’s observation of its environment, is, besides eventual preset strategies, the basis for the scheduling of learning cycles within their appropriate time intervals. Given that the model construction that is at the core of eager learning can be a time-consuming operation, the design of the critic therefore determines the computation resources that the agent exerts for its strategic knowledge management. Particular cases with regard to the circumstances of actual agent deployment may thereby demand custom-tailored answers to the scheduling of learning cycles. For exemplification, consider two functionally identical transport management agents. Let the first be deployed on an agent platform hosted by a high-capacity server cluster, while the second is deployed on an embedded telematics system attached to its managed truck. In this setting, it is evident that the latter agent needs to economize with its resources while the former needs not fear resource bounds. The problem generator component is, as Russell and Norvig highlight, responsible to suggest to the performance element such actions or courses of action that will lead to new and informative experiences. Hence, it constitutes a means by which the performance element is biased such that it also gives room to the exploration of the task environment, rather than only exploit existing background knowledge and models. When the agent exerts primary logistics activities that are measured by
17
Knowledge Management for Agent-Based Control Under Temporal Bounds
239
fixed performance standards, it is rendered evident that exploration cannot be prescribed blindly. Instead, it rather needs to be conditioned on a positive assessment that in the individual case, the agent can indeed afford to explore. This is the case, if the current situation allows for exploration that does not have a negative net effect on the quality of logistic operations. Consider, for instance, the case of a transport management agent that initiates a detour of its managed truck, knowing that such a course of action will not only constitute spatial exploration but, at the same time, will improve due-time reliability in the delivery of a certain order, which justifies additional haulage expenses as compared to the “direct” route. In other cases, the agent needs to trade short-term costs of exploration, either real monetary losses or lower quality of services, with expected mid to long-term returns on investment. This stresses, that a good situation assessment is, again, critical. With regard to the overarching theme of agent-oriented knowledge management under bounded time constraints, we can argue that the concrete design, or rather the strategies employed by, the agent’s problem generator bear influence on the time it takes an agent to accumulate a sufficiently representative pool of experience to drive the learning of high-quality decision support models. Having discussed the basic architecture of a learning agent with particular focus on design choices that influence (1) when to spend time on learning and (2) how far to stretch the gathering of suitable learning input, we focus in the following on the relevance of model compilation and maintenance along the agent life cycle. This contemplation will thereby lead yet to other issues the agent is confronted with; namely, how to enable fast adaption, especially in the face of sudden changes in the environment.
17.3.2.1 Learning Along the Agent Life Cycle We assume that the agents in the multi-agent system, on instantiation, are equipped with (1) common background knowledge about their domain of practice and (2) default models for decision support. Consider as one paradigmatic example the case of transport management agents that act on behalf of particular means of transport. Such agents may be equipped with peripheral models (1) to predict the course of operative order handling (i.e., factors such as traffic densities or severe weather conditions which directly influence transport durations, or holding times for transloading at storage facilities approached during pickup and delivery), or (2) to predict the further development of the order situation. If such models are provided, they can usually be characterized as general-purpose models. These provide a basic support for a wide spectrum of logistic tasks handled by the agent community. Broad applicability hereby takes precedence over advanced support for the special needs of particular communities of practice. As a consequence, early in their respective lifecycle, the agents are bound to operate on a basic level of performance. Consequently, they may well be thought of as novice actors. When considering the combination of a persistent multi-agent system on the one hand and a task environment that features dynamic short and long term changes, design-time default knowledge as initial
240
T. Warden et al.
support for novice agents which is furnished on system startup is bound to face an aging process whose speed is proportional to environmental change. Therefore, in order to retain the baseline performance achievable initially using provisioned default models, the latter need to be revised on a regular basis. Over time, with the continuous fulfillment of their primary logistic role, novice agents gather a body of individual experience, which constitutes the foundation for the personalization of their operative knowledge models. Local machine learning allows for the evolution of situated models that are both specifically tailored to but also invariably biased by the agent deployment context. For instance, an assessment of local learning in a reduced logistic scenario showed that the compilation of a peripheral model of the environmental influence on forwarding times on available transport relations led to increased precision in the estimation of transport durations and, as an anticipated consequence, better due-time reliability [4]. Such positive results, however, went along with several individual hardships, which needed to be handled. Relying on local learning alone usually requires a significant amount of time for any new agent to promote itself from its initial novice status to competent actor and, in some cases, expert in its community of practice.
17.3.2.2 A Closer Look at Eager Learning Figure 17.4 describes the typical steps, which need to be conducted by the learning element of an agent when employing an eager learning approach. That is, the pursued goal of the learning process is the initial construction, or the refinement and/or adaption of an existing model that generalizes from particular experience acquired by the agent. The steps in the learning process can be described by analogy to the process for knowledge discovery in large databases (KDD, cf. [3]), albeit with several notable changes. At the beginning of the learning process stands the belief base into which the agent thus far continuously integrated its observations of
Primary Logistics Role Reason
Act
Auxiliary Sustainable Knowledge Management Role Decision (Support) Models
Model Integration
Percept
Belief Integration
Environment
Experience Selection
Belief Base
Preprocessing
Target Beliefs
Transformation
Preproccessed Beliefs
Eager Learning
Transformed Beliefs
Evaluation & Testing
Model Candidate
Proofed Model
Fig. 17.4 The steps of an eager learning sweep initiated by the agent in order to construct or adapt a model (Schema adapted from [3])
17
Knowledge Management for Agent-Based Control Under Temporal Bounds
241
relevant excerpts of its situation context (Belief Integration). The agent belief base is the analogue of the data bases considered in KDD. The difference here is that all the experience data that has been memorized, was acquired on the fly during the fulfillment of the agent’s primary logistic role. When a new sweep of eager learning is scheduled, it is necessary to make a choice which subset of the belief base should be considered for learning (Experience Selection). This choice has several dimensions. The first dimension is the absolute number of records that should be used as input to the learner. The strategic options, which are conceivable in guiding the selection, are manifold and bring along characteristic follow-up challenges. For instance, the agent could opt not to make a selection at all for fear of discarding relevant data prematurely. As a consequence, successive learning sweeps would need to handle an ever-growing database as input, which is eventually bound to slow down or even inhibit learning late in the agent life cycle. Alternatively, the agent could restrict the data in different ways, amongst others by a restriction based on the absolute number of records, or the selection of records from a particular time window in the recent past, or the selection of what is considered a good representative sampling of records from the belief base. The latter strategies are thereby conceivable in fixed and adaptive form. The actual choice of the selection strategy has an impact on the maintenance of the agent belief base; in particular, for how long experience data should be retained and when it can be consequently safely forgotten since the gist of this experience has been captured in learned models, such that is can still be recollected as needed. A second dimension, which needs to be considered in the selection of experience data, is the weighting of respective records. Once the experience selection process is finalized, the next two steps, namely Preprocessing and Transformation, bring the data in a format which is suitable as input for the learning approach that is employed by the agent. In comparison with the KDD case, these process steps have subordinated importance. This is due to the fact that data gathering and learning entity are one and the same such that problems of data conversion or data processing with regard to missing or incorrect values are much less pronounced. The next step in the process chain of an eager learning sweep is then the application of the appropriate machine learning method itself. The result of this step is a new model candidate which might be suitable as a substitute of the respective model that is used thus far in for the execution of the primary agent roles, be it a default model or a self-learned model from a previous learning sweep. In order to assess whether the new model candidate is indeed superior to existing models within the momentary agent deployment context, the next step in the process chain is the model evaluation. To this end, the performance of the model, for instance in terms of precision and recall, needs to be assessed based on benchmark cases that are representative for the respective agent. These benchmark cases can be sampled directly from the agent’s own pool of experience. That is, while the eager learning approach can function as a means for the agent to forget aged, and especially repetitive experiences, it must still be ensured that a representative benchmark set is retained. The final step within a successful learning sweep, that is where the model candidate
242
T. Warden et al.
could be affirmed in the evaluation step, is then the operationalization of the new model. Open questions with regard to this step concern the question, how many models the agent should keep for a possible future reassessment and consequently, a possible reinstatement.
17.3.3 Employment of Models in Autonomous Decision-Making Having cast different learning approaches and their prerequisites in their temporal context for the case of autonomous logistic multi-agent systems, we will now shift our focus to the employment of the models in decision-making. Knowledge acquired for autonomous decision-making is encoded in models for different, dedicated purposes. One can distinguish primary models for decision making on the one hand, such as Markov decision processes with utility functions, and secondary decision support models on the other hand, which provide helpful insights to the decision situation at hand. In the following we will focus on models for decision support. They provide knowledge and inference consulting the agent in terms of situation analysis as well as state estimation and prediction in uncertain and partially observable environments. Presupposing that these models are available and support informed deliberate decisions in principle, there is still a need for an adequate selection and usage. This can be seen in three perspectives: (1) model applicability and quality, (2) inference, and (3) situation awareness.
17.3.3.1 Model Applicability and Quality The first perspective in this matter is the applicability and quality of a considered model in the current decision situation. That is, the decision context must fit the domain (as class of contexts) the employed model was designed for. For instance, a traffic prediction model might only be applicable or reliable in a specific region or even only a specific day of the week. If a model covers a larger domain it is also desirable that parameters and observations of a specific situation are reflected by discriminating variables in the model. In order to enable the agent to assess the applicability of available models, each model must provide meta-data about its application domain (i.e., a domain explication). On the other hand the agent needs context explication to match its situation with the domain of the model [18, 19]. The quality aspect considers statistical significance and information entropy of a generated model. The data underlying the model knowledge in terms of generalized data must be representative and meaningful for the intended domain. This is opposed to a model that originates form (sparse) data obtained by chance. Consequently, this second issue relates to the time needed to obtain enough information for generating a representative domain approximation. Thus, a model needs statistical grounding information (e.g. number of observations underlying the model or confidence
17
Knowledge Management for Agent-Based Control Under Temporal Bounds
243
intervals) for its quality and reliability to be assessed for the agent’s current decision situation.
17.3.3.2 Inference The second perspective, focused here, concerns the inference when applying a model. Inference is the process of applying a model to extract new knowledge that was not explicitly known before. Here we will focus on deduction, i.e., concluding a specific fact from more general knowledge (e.g. by a rule), and probabilistic reasoning, i.e., computing the probability of an uncertain event from statistical knowledge (e.g. encoded in a Bayesian network) in presence of certain evidence. In particular one must consider the time consumed during inference and the precision of that inference. While one intends to have precise results in shortest time, both factors are conflicting in general. Consequently, a usual design trade-off is between representation expressivity and inference efficiency when using a representation. Indeed, some model representations, such as first order logic, cannot even guarantee that an inference will yield a result at all. In fact, the model will not know that it cannot find a result and thus might compute forever. Thus, such representations might shape up as being even ineffective and logical expressivity should be limited to decidable subsets of first order logic, e.g. description logics applied in ontologies. At runtime there is also a trade-off between time and precision when applying approximate inference. In particular this holds for probabilistic inference where there is stochastic sampling as a prevalent approach for estimating probabilities from representative examples drawn based on the statistic knowledge encoded in the model [7,23]. The resulting estimated probability is supposed to converge to the exact probability with an increasing number of samples. However, the more samples are computed the longer inference will take. With increasing complexity of probabilistic models approximate inference will vanish as a choice and become a necessity for efficient reasoning. Thus, the agent will need to choose an appropriate amount of samples on the one hand and handle imprecise results on the other hand. An adequate choice for the amount of samples is closely connected to a good estimation for the time and computational resources available until the result of inference is needed for a decision. Sampling would also allow for an iterative approach where the inference result is enhanced by more samples until there is no time left thereby making it an anytime approach. But when decisions rely on the result, this kind of procedure is risky as long as there is no estimation of gained precision. In fact, it could be better to postpone a decision until inference is precise enough. Thus, available time does not need to be considered as a fixed number in every case. When reasoning about inference efficiency, one also has to bear in mind that approximate inference is not always better. Preprocessing of probabilistic models can also allow for fast but still exact inference as long as stochastic dependencies are not too ramified (cf. [15]).
244
T. Warden et al.
17.3.3.3 Situation Awareness The third perspective concerns situation awareness, i.e., the amount of relevant situational knowledge or information available to the agent (cf. [5]). In dynamic and spacious application domains the agent faces partial observability of its environment and thus needs to cope with domain-inherent uncertainty about the state of affairs. In probabilistic networks the available information corresponds to evidence and allows for more situation-specific state estimations or predictions. When using logical representations incomplete information may even lead to wrong inferences (cf. [5]). This may result from inference mechanisms using, e.g., negation-as-failure, which assume everything unknown and not otherwise inferable as being false. Regarding time constraints the issue of situation awareness is intertwined with information acquisition for gaining evidence. That is, the agent needs to gather environment information (e.g., via own sensors) in order to better assess the situation it is in and draw corresponding conclusion for the best course of action. Because there might be a large amount of possible information and the resources for acquiring them are limited, the agent needs to focus on the most relevant information for the decision at hand. The challenge here is to find an adequate measure for information relevance and to identify and rank useful information based on that measure. Additionally this has to be done quickly and thus without assessing every potential information. With regard to relevance measures Howard (1988, 1990) developed a theory of information value, which quantifies relevance based on influence on expected utility of a decision [8,9]. That is, the value of a potential observation is given by the difference in expected utility when considering the observation’s outcome or not. Because the observation has not been made yet, its influence is estimated by aver-aging the influence of each possible observation outcome weighted by its probability. Usually decision networks (also known as influence diagrams) are employed as a decision problem representation when computing information values. While Howard’s relevance measure is sound with respect to decision theory, its computation is very time-consuming unfortunately. The determination of information values is proven to belong to the #P complexity class (i.e., counting problems associated with the decision problems in NP) even for very simple models [12]. Thus, approximate algorithms have to be considered again and information value computations should be limited to the smallest possible subset of potential observations. Fortunately, the graph-theoretic property of d-separation [17] in Bayesian and decision networks (i.e. structure-derived stochastic independence) allows identifying a subset of nodes that is guaranteed to include the most valuable single information and maximizes the total possible in-formation value of the network. This subset is given by all chance nodes that are not d-separated from a reference utility node by any other single chance node together with current evidence [6]. For all nodes within this subset the information value has to be computed in order to identify the one that provides the most valuable information. On this basis a goal-oriented algorithm for information acquisition can be de-signed which incrementally increases situation awareness by acquiring
17
Knowledge Management for Agent-Based Control Under Temporal Bounds
245
additional information in order of relevance [5, 6]. When this algorithm is executed in parallel to the primary decision process it can run until, eventually, a decision has to be rendered. Because such information acquisition improves expected utility and can be interrupted at any point without harm it can be considered an anytime approach.
17.4 Conclusions and Outlook In this work we have addressed different types of activities of adaptive agents in autonomous logistics. For this, we differentiated between primary and auxiliary agent roles and associated activities. As both types of activities require careful planning and scheduling, we introduced a description framework that allows explicating temporal ordering and duration constraints of agent tasks. We have extended this description model to individual knowledge management sub tasks that are relevant, both in the context of a concrete decision situation and continuous adaption and maintenance of decision support models, such as information acquisition, model formation and model exploitation in autonomous decision-making. In doing so, we have identified different types of trade-offs that need to be considered at design time and at run time of an agent that is part of a multiagent-based logistic control system. As we have shown above, the autonomous control of logistic processes by means of multi-agent systems requires solving individual knowledge-intensive tasks, that each come with their respective temporal price. The concrete temporal price tag associated with the individual knowledge management tasks has to be considered accordingly in the design and at run-time of an autonomous logistic multi-agent system. In future work, we seek on the one hand to flesh out the introduced description framework and relate it to approaches in automated planning and scheduling. On the other hand, concrete strategies for an intelligent scheduling of knowledge management tasks need to be developed and consequently be assessed using, for instance, multiagent-based simulation. A particular focus of an evaluation thereby should be the correlation of strategies for effective, context-sensitive knowledge management and the respective logistic performance that is measured by an application-specific system of key performance indicators.
References 1. Allen JF (1984) Towards a general theory of action and time. Artif Intell 23(2):123–154 2. Bloos M, Schönberger J, Kopfer H (2009) Supporting cooperative demand fulfillment in supply networks using autonomous control and multi-agent-systems. In: INFORMATIK 2009, pp 3590–3604 3. Fayyad U, Piatetsky-Shapiro G, Smyth P (1996) The KDD process for extracting useful knowledge from volumes of data. Commun ACM 39(1):27–34
246
T. Warden et al.
4. Gehrke JD, Wojtusiak J (2008) A natural induction approach to traffic prediction for autonomous agent-based vehicle route planning. Reports of the machine learning and inference laboratory, MLI 08–1, George Mason University, Fairfax, VA 5. Gehrke JD (2009) Evaluating situation awareness of autonomous systems. In Madhavan R, Tunstel E, Messina E (eds) Performance evaluation and benchmarking of intelligent systems. Springer, pp 93–111 6. Gehrke JD Informierte entscheidungsfindung autonomer systeme in dynamischen umgebungen. Doctoral dissertation, Universität Bremen, Germany (unpublished) 7. Henrion M (1986) Propagating uncertainty in Bayesian networks by probabilistic logic sampling. In: Proceedings of the 2nd annual conference on uncertainty in artificial intelligence (UAI-86). Elsevier, pp 149–164 8. Howard RA (1988) Information value theory. IEEE Trans Syst Sci Cybernet SSC 2(1):22–26 9. Howard RA (1990) From influence to relevance to knowledge. In Oliver RM, Smith JQ (eds.) Influence diagrams, belief nets and decision analysis. Wiley, pp 3–23 10. Hribernik K, Warden T, Thoben KD, Herzog O (2010) An internet of things for transport logistics – an approach to connecting the information and material flows in autonomous cooperating logistics processes. In: Proceedings of the 12th international MITIP conference on information technology & innovation processes of the enterprises, pp 54–67 11. Jedermann R, Gehrke JD, Becker M, Behrens C, Morales Kluge, E, Herzog O, Lang, W (2007) Transport scenario for the intelligent container. In: Hülsmann M, Windt K (eds) Understanding autonomous cooperation & control in logistics. Springer, pp 393–404 12. Krause A, Guestrin C (2009) Optimal value of information in graphical models. J Artif Intell Res 35(1):557–591 13. Kirn S, Herzog O, Lockemann P, Spaniol O (2006) (eds) Multi-agent engineering theory and applications in enterprises. International handbooks on information systems. Springer 14. Langer H, Gehrke JD, Herzog O (2007) Distributed knowledge management in dynamic environments. In: Hülsmann M, Windt, K (eds) Understanding autonomous cooperation and control in logistics. Springer, pp 215–231 15. Lauritzen SL, Spiegelhalter DJ (1988) Local computations with probabilities on graphical structures and their applications to expert systems. J Royal Stat SocB 50(2):157–194 16. Mitchel T (1997) Machine learning. McGraw-Hill 17. Pearl J (1988) Probabilistic reasoning in intelligent systems: networks of plausible inference, 2nd edn. Morgan Kaufmann 18. Porzel R (2010) Contextual computing: models and applications. Cognitive Technologies, Springer, Heidelberg, Germany 19. Porzel R, Warden T (2010) Working simulations with a foundational ontology. In: Schill K, Scholz-Reiter B, Frommberger L (eds) Proceedings of the workshop on artificial intelligence and logistics at the 19th European conference on artificial intelligence, Lisbon 20. Russell S, Norvig P (2003) Artificial intelligence – a modern approach, 2nd edn. Prentice Hall 21. Schuldt A (2010) Multi-agent coordination enabling autonomous logistics. Doctoral dissertation, Universität Bremen, Germany 22. Warden T, Porzel R, Gehrke JD, Herzog O, Langer H, Malaka R (2010) Towards ontologybased multi-agent simulations: The PlaSMA approach. In: 24th European conference on modelling and simulation (ECMS 2010). European council for modelling and simulation, pp 50–56 23. Yuan C, Druzdzel MJ (2007) Generalized evidence pre-propagated importance sampling for hybrid Bayesian networks. In: Proceedings of the 22nd AAAI conference on artificial intelligence, pp 1296–1302. AAAI Press
Chapter 18
Impacts of Data Integration Approaches on the Limitations of Autonomous Cooperating Logistics Processes Karl A. Hribernik, Christoph Kramer, Carl Hans, and Klaus-Dieter Thoben
18.1 Introduction Autonomous cooperating logistics processes are characterized by the ability of intelligent logistics objects to process information, to render and to execute decisions on their own. In order to do so, these objects need to be able to access data relevant to the decisions to be made. For example, an intelligent container, which “wants” to be transported from A to B, needs to interact with many different entities, such as different sensor networks, freight forwarders or cold storages. This means that, in order for an IT infrastructure to truly support autonomous cooperating logistics processes on an operational level, its intelligent logistics objects not only need to be able to communicate with each other, but also be suitably integrated into the overall logistics IT landscape. The “traditional” IT landscape in logistics is already a highly complex, distributed and heterogeneous one even without taking autonomous cooperating processes into account. Significant effort was and still is spent in order to achieve at least integration between systems of certain business partners by bridging the technological islands through specific ICT solutions [15]. However, most of these solutions sooner or later become obsolete due to both the continuous development of individual standards and systems and the highly dynamic partnerships found in today’s enterprise networks. Instead of developing solutions for 1:1 relationships, it would be preferable to develop a general solution which allows a unique access to all relevant logistics data while accepting the diversity of existing systems and standards [16]. This situation is exacerbated by developments in modern logistics such as autonomous cooperating logistics processes and the Internet of Things. Both developments lead to the creation of “new islands” of technology development in the IT logistics landscape. Depending on the application, relevant data may be stored in heterogeneous enterprise systems, such as Warehouse Management K.A. Hribernik (B), C. Kramer, C. Hans, and K.-D. Thoben Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany e-mail:
[email protected] M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_18,
247
248
K.A. Hribernik et al.
Systems (WMS), Enterprise Resource Planning Systems (ERP) or disposition systems. At the same time, data from item-level tracking and tracing systems needs to be taken into account, in particular that pertaining to RFID. Data may also be generated and stored in systems embedded into logistics objects such as trucks or containers, or be generated dynamically, for example by sensor networks monitoring the temperature of a refrigerated container. Whilst the specific requirements towards data integration differs according to the characteristics of each individual application of autonomous control, it can be said that, in general, digital counterparts representing individual logistics entities need to be able to access data relevant to their decision making processes, regardless which “island” that data may be located on. The heterogeneity of the data sources, their highly distributed nature, along with their availability makes the selection of an adequate data integration mechanism a highly challenging task. Furthermore, research in autonomous cooperating logistics processes shows that different control problems arise from different applications of autonomous control, resulting in a wide spectrum of degrees of autonomy [42]. The resulting requirements towards the characteristics of intelligent logistics objects involved in these processes as well as the underlying data processing, decision making and consequently data integration strategies vary in accordance with the degree of autonomy. When considering an adequate approach to data integration for the support of autonomous cooperating logistics processes, this needs to be taken into account along with the characteristics of the underlying IT landscape outlined above. Numerous approaches to data integration exist which may be taken into consideration. Each approach exhibits a number of strengths and weaknesses which characterise its applicability to the problem of data integration for the support of autonomous cooperating logistics processes. It follows that these characteristics impact on the autonomous cooperating logistics processes themselves. For example, a data integration approach which is weak in providing real-time access to heterogeneous data sources might negatively impact the reactivity of the overall autonomous cooperating logistics system. That means the selection, configuration and implementation of a data integration approach has a direct effect upon the degree and limits of the autonomous cooperating logistics system it is chosen to support. Aspects such as the timeliness, reactivity, scalability, robustness and adaptability of the data integration mechanisms employed factor into limiting the degree of possible autonomous cooperation. This contribution endeavours to systematically analyse the limitations thus imposed by different data integration approaches, and to establish a scheme of categorization for identifying adequate data integration approaches for different degrees of autonomous cooperation. It is structured as follows: first, the theoretical background of the problem area is discussed. This encompasses autonomous cooperating logistics processes, the types of data source involved in such processes, and an overview of data integration approaches. The next section deals with the analysis and categorisation of different data integration approaches. An approach to judging the effect the different categories of data integration approach will have on different dimensions of autonomous cooperating logistics processes is derived.
18
Impacts of Data Integration Approaches
249
The next section applies the approach to the different categories of data integration in order to present a comparison of the different approaches to data integration and their effect on autonomous cooperating logistics processes. A summary and outlook concludes this contribution.
18.2 Theoretical Background The following sections present an overview of the theoretical background relevant to this contribution. First, autonomous cooperating logistics processes are introduced. Then, an overview of the IT landscape in those processes illustrates the integration targets required to be handled by potential data integration mechanisms.
18.2.1 Autonomous Cooperating Logistics Processes In the context of this contribution, the term “Autonomous Control” is used following Böse and Windt (2007) to describe: “. . . processes of decentralised decision-making in heterarchical structures. It presumes interacting elements in non-deterministic systems, which possess the capability and possibility to render decisions independently.” [4]
The research area of autonomous cooperating logistics processes [27] aims to meet today’s logistics challenges such as the goods structure, logistics and structural effects identified by Aberle (2003), by introducing autonomy and self-organisation into control, information processing and decision-making in logistics [1, 10]. The argumentation is that central control and planning of logistics processes has reached its limits in addressing these issues [27]. Here, the term “autonomy” describes: “. . . the capability of a system, process or an item to design its input-, throughput- and output-profiles as an anticipative or reactive answer to changing constraints of environmental parameters.”
The application of autonomous control to logistics processes is expected to increase their robustness, flexibility, adaptability and reactivity to respond to changing business environments, requirements and to changing or partially conflicting objectives [27]. A prominent characteristic of this understanding is the decentralisation of decision-making responsibilities in contrast to traditional, hierarchical process control. A dynamic heterarchy in which otherwise passive logistics entities are equipped with the ability to process information, to render and execute decisions on their own replaces the strict centralised top-down management of traditional logistics processes. Artificial agents are entrusted to act in their own “best interest” within the bounds of their operational, tactical or strategic [34] autonomies. The motivation for this approach is, amongst others, an expected improved robustness and increased scalability of process control.
250
K.A. Hribernik et al.
The concept of an intelligent logistics object is inherent in the understanding of autonomous control in logistics systems [4]. Here, “. . . autonomous control in logistics systems is characterized by the ability of logistics objects to process information, to render and to execute decisions on their own.”
Logistics objects are defined in this context as both, “. . . material items (e.g. parts, machines or conveyors) and immaterial items (e.g. production orders) of a networked logistics system, which have the ability to interact with other logistics objects of the considered system.”
In Scholz-Reiter et al. (2007), the former are further differentiated as commodities and all types of resources whilst constraining the immaterial logistics objects to orders [28]. According to this understanding, an intelligent logistics object is consequently either a material or immaterial logistics object which is capable of communicating and interacting with other logistics objects. It is a broader understanding than that of the Internet of Things which additionally encompasses autonomous objects without physical representations.
18.2.2 The IT Landscape in Autonomous Cooperating Logistics Processes Hribernik et al. (2010) categorises the major data sources which comprise the potential IT landscapes supporting autonomous cooperating logistics processes (cf. Table 18.1) [18]. Here, four types of data source are differentiated: 1. Logistics IT systems, describing IT systems in logistics such as ERP, WMS, disposition and other “traditional” enterprise systems used in logistics. 2. Intelligent material logistics objects – which relate to material intelligent logistics objects, which exhibit characteristics of the PEID (Product Embedded Information Device) classification scheme [33]. 3. Digital counterparts – these relate to the decision making components of intelligent logistics objects, whether located in the object or in the network. 4. Sensors and actuators – relating to sensors, sensor networks and actuators, which fall outside of the previous categories. With regards to logistics IT systems, EDIFACT EANCOM and SAP RFC are the most prominent targets. However, the more than 30% systems with proprietary interfaces cannot be neglected [18]. Consequently, a data integration approach must be able to cope with both semi-structured, standard data exchange formats as well as function interfaces and be flexible enough to cope with arbitrary proprietary interfaces. To integrate intelligent material logistics objects, the support of RFID middleware standards such as the EPCglobal Framework Architecture, foremost EPCIS, is
18
Impacts of Data Integration Approaches
251
Table 18.1 Major data sources supporting autonomous cooperating logistics processes Data sources Type(s) Interface/standard Logistics IT systems General EDIFACT EANCOM EANCOM XML ebXML SAP compliant SAP RFC (Remote Function Call) Other Bespoke proprietary Intelligent logistics objects EPC compliant EPCIS ID@URI compliant Dialog PEIDs PMI OSGi-based OSGi Other Bespoke proprietary Digital counterparts Multi-agent based (e.g. JADE, ACL (Agent Communication PlaSMa, Dialog) Language) Agent proxies Dialog agent EDIFACT EANCOM Sensors & actuators Java-based OSGi OGC compliant SensorML PEIDs PMI Other sensors Bespoke proprietary formats OPC OPC DA OPC XML DA OPC AU General GDI ORiN API Smart Embedded Devices in SOCRADES Manufacturing Other actuators Bespoke proprietary formats
mandatory [12, 30, 35]. In addition, a means to interfacing emerging standards for the integration of PEIDs and other embedded devices is necessary. The PROMISE Messaging Interface PMI currently offers the most comprehensive and structured approach to this. The field of digital counterparts is dominated by software agent technology. The PlaSMa platform is dedicated to the support of autonomous cooperating logistics processes and is consequently of highest priority. Other approaches favour service interfaces. The possibility of agent communication via EANCOM strengthens the need for EANCOM support, but is at the present time not widespread. Sensor and sensor network integration is at the present time largely a case-bycase decision, with most interface using proprietary approaches. However, emerging standards such as PMI or SensorML are increasing in importance and should not be neglected. A data integration approach therefore needs to be highly flexible towards sensor data sources. With regards to actuators, a promising contribution can be found in the Unified Architecture standards put forwards by OPC. A proposed data
252
K.A. Hribernik et al.
integration approach should also take into account the standards emanating from ISO 20242 and factory automation initiatives such as SOCDRADES.
18.3 Categories for Data Integration Approaches Various approaches for the integration of heterogeneous data sources exist. They all offer not only consistent access to data but also the ability to resolve existing integration conflicts. Literature suggests different categorisation schemata for the classification of these approaches. One scheme differentiates between the different information system architecture layers: manual integration, common user interface, integration by application, integration by middleware, uniform data access and common data storage [44]. In the context of autonomous logistics processes, both manual integration and common user interface can be disregarded – the data to be integrated is consumed not only by human users but primarily by other system components, including enterprise systems and distributed decision making components, e.g. agents in intelligent logistics objects. The remaining four categories describe different levels of ways of coupling information systems along a continuum moving from placing full responsibility for data integration with the querying application (integration by application) through to accomplishing logical integration at the data access level (uniform data access). The final category describes transferring all data to be integrated into a new data storage system (common data storage). Another widely accepted categorization scheme also focuses on the type of coupling between integrated data sources [36]. Here, tightly coupled, loosely-coupled and object-oriented approaches are distinguished. Tightly-coupled approaches correspond roughly to above category uniform data and include common data storage access, whilst loosely-coupled approaches find their place at the opposite end of the continuum. The category “object-oriented” introduces a further level of detail into the classification of data integration approaches. It is important to understand that these are not absolutes – the properties of a data integration mechanism might signify it belongs to more than one category. In the following, the three categories are described in more detail.
18.3.1 Tightly-Coupled Approaches An integration mechanism is considered tightly-coupled if it is based on one or more federated schemata which solve the integration conflicts. The federated schemata constitute a consolidation of the local schemata of the integrated database systems. An integration approach which uses federated schemata is called a “federated database management system”. Applications that want to access data from the individual subsystems of a federated database management system interact only with the federated schema. The
18
Impacts of Data Integration Approaches
253
various schemata of the subsystems remain hidden from the application. Direct access is not possible. If a request is made to the federated database management system, it is responsible for breaking down the request into component queries that correspond to the respective local schemata of the component systems. Subsequently, the data supplied by the subsystems is assembled. In developing a federated schema, care needs to be taken to ensure it results in a superset of the local schemata. If this is not the case, data is lost because it can’t be retrieved by the applications. Creating a federated schema is a challenging task in the development of closely coupled systems [32]. Semantically equivalent data has to be identified and the resulting integration conflicts have to be resolved. The effort for resolving the various conflicts of integration varies considerably. This allows resolving syntactical problems, such as data type conflicts, relatively quickly and easy. Other conflicts however, are difficult and expensive solvable.
18.3.2 Loosely Coupled Approaches Loosely coupled integration mechanisms use query formalisms to enable applications to define the mappings to data sources themselves. Such a query formalism needs to support a multi-database query language. The query language must, in addition to its ability to query different data sources simultaneously, allow the definition of integration rules which are able to solve the different integration problems. This means that a loosely coupled integration mechanism does not constitute a readymade solution for all integration problems, but only provides the means to solve such problems in the form of a complex query language. This means the integration mechanism isn’t responsible for solving the integration conflicts – this responsibility is passed on to the querying application.
18.3.3 Object-Oriented Approaches Object-oriented integration mechanisms are very similar to tightly ones. Both exhibit global schemata for the elimination of the integration conflicts. In objectoriented approaches, data is encapsulated in objects. Semantically equivalent objects from different data sources are combined into the federated schemata from a super type. Unlike traditional object orientation, functions and methods of each object are inherited backwards. Thus, the super type inherits all the functionality of collected objects. Functions are created within the super type which access the different fields and functions of the collected objects and return them. The integration conflicts are resolved in these functions. An overview about different object-oriented approaches is given by [24].
254
K.A. Hribernik et al.
18.4 Classification of Data Integration Mechanisms The following sections present a number of major data integration mechanisms, which may be classified according to the scheme outlined in the previous section. Some mechanisms cannot clearly be classified. In order to reflect this, Fig. 18.1 presents a sketch showing how roughly the major data integration mechanisms relate to the three categories at the corners. Subsequently, a classification of the mechanisms is shown according to the categories outlined in the previous section. The inclusion of Message Oriented Middleware, Service-Oriented-Architecture and Enterprise Service Bus is motivated by the classification according to [44] and represent widely adopted approaches for the facilitation of integration by application.
18.4.1 Data Warehouses A data warehouse is a centralized data pool which mirrors data provided by heterogeneous and often distributed data sources in a local database with a previously defined schema [37]. Data warehouses are often used for the Online Analytical Processing (OLAP) or as basis for subsequent data mining [20]. A data warehouse usually has three main functions: data extraction and updating, data integration and data storage. The data warehouse extracts the required
Object oriented
Object oriented Database
Semantic Mediator Federated Database System Data warehouse
Tightly coupled
ODS
Fig. 18.1 Classification of data integration mechanisms
PDMS
MOM
SOA ESB
Loosely Coupled
18
Impacts of Data Integration Approaches
255
information from the various heterogeneous data sources. This extraction is executed periodically in a preset interval in order to keep the data up to date. Accordingly, the extracted information is transformed and integrated into the local schema provided by the data warehouse. This local schema and the respective database tables have been initially developed at the creation of the data warehouse. For integration and transformation, a mapping between the local schema and the schema of the respective data source is needed. By means of this mapping the different integration problems are solved. Following to the integration, the data is stored in a local database. Afterwards, the integrated information is accessible by users through this local database. With the use of a local schema, data warehouses falls into the category of tightly-coupled integration approaches.
18.4.2 Operational Data Stores (ODS) This kind of integration approach is very similar to the data warehouse approach [19]. Just as a data warehouse, the operational data store extracts data from the various data sources to be integrated. The extracted data is transformed and integrated towards the global schema and stored in an own local database. The main difference between operational data stores and data warehouses is the fact, that usually in operational data stores only current data sets are stored, while in data warehouses historical data is stored also. Whereas a data warehouse would store data from last 50 days for example, an ODS would only store the most current data [3]. So, the ODS stores only data useful for the current operational world.
18.4.3 Federated Database Systems (FDBS) Federated Database-Systems are systems that offer access to autonomous and often heterogeneous data sources [29]. Unlike data warehouses or operational data stores, the FDBS don’t mirror all the data from the various data sources in an own database. Instead, in case of a query, the required data sets are queried from the data sources, integrated and preprocessed at runtime. A FDBS offers a federated global schema which is a combination of the local schemata from the data sources [17] and which was build at construction time manually. There are two main approaches for creating a federated schema. In the globalas-view-approach, the relations of the global schema are defined as view over the relations of the local schemata. In contrast, the local-as-view-approach defines relations of the local schemata as view over the relations of the global schema. There are also approaches which are combinations of the local-as-view and global-as-view approaches like the one introduced by [43]. These approaches try to combine the various characteristics of the two approaches in order to meet their needs.
256
K.A. Hribernik et al.
18.4.4 Message Oriented Middleware (MOM) In Message-oriented Middleware environments, communication participants interchange data via messages in an asynchronous way. Thereby the messages aren’t exchanged directly between the participants; instead the communication takes place via a middleware. The middleware receives messages from the sender and passes them to the receiver. Thus, an asynchronous communication between sender and receiver is possible. The used middleware is a kind of abstraction layer between the various communication partners. Since the single sender and receiver just interact with the middleware, they don’t need to know many details about each other; the message could be transformed by the MOM to fit the requirements of the receiver [21]. Thus, the participants in a system are loosely coupled [7]. Since the messages send over a MOM are normally used to transfer information and data, the MOM systems represent a form of data integration.
18.4.5 Service-Oriented-Architecture (SOA) In service oriented architecture, applications offer their functionality as reusable services. These services implement a standardized and implementation independent interface. A service in a SOA represents a complete self-contained business function [22]. Existing services may be reused by applications or by other services [11], therefore interoperable protocols exist. In order to share knowledge about services and their existence, a SOA normally makes use of a naming service. All available services are registered by their providing application to a central naming service with a service description. Consuming applications can discover the available services with this naming service and query the service description to obtain the knowledge about the functionality and how to access the service. Combining different services to implement a complex business process is called service orchestration and offers one of the main goals of business integration.
18.4.6 Enterprise Service Bus (ESB) Out of the message-oriented middleware and service-oriented architecture the concept of the enterprise service bus (ESB) was evolved. The concept of ESB tries to solve the disadvantages of MOM and SOA such as increasing point-to-point interfaces through services in a SOA [8] or the use of proprietary protocols and platform specific interfaces in MOM solutions [21]. An ESB is an open-standards based integration infrastructure for applications or services. It provides a messaged-based communication between distributed applications or services in a secure and trustworthy manner. Therefore, the a ESB System
18
Impacts of Data Integration Approaches
257
provides transformation of messages to fit the requirements of the services, routing, message acceptance, message processing and message routing as well as sending messages to multiple receivers [23]. ESB are usually realised out of three elements: A message broker (similar to MOM) or a Message protocol, adapters or service container and mediation/ integration services. Services are connected to the ESB by using adapters and communicate over the message broker. The elements of the ESB which deals with the previous mentioned features of the ESB are implemented and attached to the ESB as services also.
18.4.7 Object-Oriented Databases In object-oriented databases the data is stored in an object-oriented manner, similar to object-orientation in programming languages. Thus it supports an objectoriented data model instead of a relational data model like traditional relational databases. The object-oriented data model provides concepts like encapsulation, classes, inheritance, overriding and object identification [2]. Multidatabase Systems are an approach for data integration where a single module is located on top of the data sources to be integrated. This component offers a global schema to the user. A restriction of this approach is that it is only applicable to data sources with database management system capabilities. With the introduction of object orientation to multidatabase systems, this restriction has been removed [9]. With the concepts of object-orientation it’s possible to encapsulate every possible data source with the use of implanted wrapper-components.
18.4.8 Peer-Data-Management Systems In Peer-Data-Management Systems there is no central intermediate layer. Instead the data sources are encapsulated by peers. Peers interact as autonomous unit which are able to answer queries. Each peer may be connected through mappings to other peers. To answer a query, the peer use the data of the encapsulated data source as well as data from peers with which he is connected through mappings. According to [26] and [32], a peer consist of a peer schema, a set of local schemata, local mappings and peer mappings. The peer schema issues what data the peer provides, it can be denoted as export schema. The local mappings connect the peer schema with the local schemata of the local data source. The peer mappings defines the relationships between the different peers in the peer-data-management system.
258
K.A. Hribernik et al.
18.4.9 Semantic Mediators The concept of mediator according to [39] provides the mediator as an independent intermediate layer between the applications which want to query data and the associated data sources. The applications of this concept put their questions to the mediator instead of being sent directly to the data sources. The mediator will forward those requests to the data sources, evaluates their returns and then generates the result for the question of the application. To fulfill this role, the mediator is composed of, according to [6] the mediator itself and the wrappers. The wrappers provide a uniform access to heterogeneous data sources. They have the knowledge that is required for access the particular data source and thus represent an abstraction of the individual characteristics of the data sources. Generally, one wrapper is necessary per heterogeneous data source. If the mediator receives a request for the schema provided by him, it splits the request and forwards the request to the wrappers. Only these parts of the request will be forwarded to the wrappers that can be answered by the appropriate data sources. The different wrappers of a mediator receive the requests of the mediator in a uniform query language, transform this request to the appropriate query language of the data source and send that request to the data source. The result of this request is then converted into a common format and returned to the mediator by the wrapper [14]. Then the mediator is responsible for integrating the different results of each wrapper and passing them to the requesting application. In semantic mediators, both syntactic and semantic descriptions of the data to be integrated are applied. The semantic mediator is capable of extracting knowledge regarding the data structures of the underlying data sources and subsequently transforming, decomposing and recomposing data requests according to that knowledge. Given a user query, the mediator first decides which data sources are responsible for the query, based on the semantic descriptions. Then, the queries for the responsible data sources are built with use of the syntactic descriptions. The results from the data sources are afterwards transformed and integrated using both, syntactic and semantic, descriptions. For the semantic description ontologies are a commonly used mean as mentioned by [38].
18.5 Identification of Limitations and Potentials A catalogue of criteria for gauging the degree of autonomous control in logistics systems is presented in [41]. Here, criteria affecting the degree of autonomous control are categorised by system layer (decision, information and execution) and each described using a number of properties. To give an example, “Location of data processing” is a criterion in the system layer “Information System.” Its properties describe a continuum between “central” and “decentralised” data processing. The
18
Impacts of Data Integration Approaches
259
further towards decentralised data processing this criterion tend in a given system, the higher the degree of autonomous control. At the present time, this catalogue of criteria encompasses three criteria for measuring the effect of the information system layer on the level of autonomous control. These are “location of data storage”, “location of data processing”, and “interaction ability”. Their level of detail is not able to capture well the impact of data integration of autonomously controlled systems. Therefore, this contribution intends to suggest additional criteria by which that impact may be captured. In order to do so, the following sections first discuss criteria of data integration mechanisms in relation to their impact on autonomous cooperating logistics systems. These criteria and their properties may then be used to extend the existing catalogue of criteria. Taken separately, they can be used to investigate the value of different data integration mechanisms for use in autonomous cooperating logistics processes.
18.5.1 Criteria of Data Integration Mechanisms for Autonomous Cooperating Logistics Processes Table 18.2 shows properties of mechanisms which provide data integration in the context of autonomous cooperating logistics processes. The properties are divided into two categories. The first deals with properties related to the reliability of a data integration mechanism, the second with its flexibility. The following sections describe in more detail both the categories and their respective properties.
18.5.1.1 Criteria of Reliability The first category is that of reliability. This refers to the capability of a data integration mechanism to perform its function correctly in a specified period of time under stated operation conditions.
Table 18.2 Properties of data integration mechanisms for autonomous cooperating logistics processes Category Criterion Description Reliability
Flexibility
Data timeliness Reactivity Robustness Quality of data Data volume scalability Data source dcalability Data source agnosticism Adaptability
Ability to guarantee up-to-date data Ability to guarantee a timely response to a query Ability to function reliably under any circumstance Ability to guarantee determinable quality of data Ability to manage increases in data volume Ability to manage increases in the amount of data sources Ability to integrate different types of data source Ability to react to changes in data sources
260
K.A. Hribernik et al.
Timeliness This criterion describes how up-to-date the data retrieved using the data integration mechanism is. A service-oriented architecture, for example, the data retrieved from a system via its service interface can be guaranteed to be valid at the time the query was accepted by the service. However, the timeliness of data retrieved from a data warehouse is dependent on the scheduling of data extraction – depending on the configuration of the system; it might be minutes or days old. In both cases, in the period of time between the query and the receipt of the data, it may have become outdated with new data. The former can be said to be timelier due to the more direct access to the data. This criterion directly impacts the behaviour of autonomous cooperating logistics processes in that a decision system may operate more dynamically and reliably the more the timeliness of the data it requests can be guaranteed.
Reactivity This criterion refers to the ability of a data integration mechanism to respond to a query within a determinable amount of time. This is dependent on the one hand on the connectivity interface used by data integration mechanism and on the other on the architecture underlying it. For example, querying a data warehouse using ODBC can be expected to generate a timely response within a determinable and short amount of time, whilst querying a Peer Data Management System might timeout without returning a complete result set from the respective peers at all.
Robustness Robustness refers to the ability to perform reliably under any circumstance. Like reliability, it is also related to scalability and timeliness as it implies the guarantee of a determinable quality of service.
Data Quality Data quality (DQ) is understood as the level of fitness for the use of the data by the data consumer in an information system. Strong et al. (1997) sub classify data quality in four Categories: intrinsic, accessibility, contextual and representational data quality [31]. Each of these categories has several dimensions as shown in Table 18.3. In autonomous logistic processes it’s necessary to perform on high quality data, to avoid defective decision making based on measurement errors in sensors for example. Hence, the quality of the provided data is an important measuring point for the various integration approaches.
18
Impacts of Data Integration Approaches
261
Table 18.3 Data quality categories and dimensions (Strong et al. 1997) DQ category DQ dimensions Intrinsic DQ Accuracy, objectivity, believability, reputation Accessibility DQ Accessibility, access security Contextual DQ Relevancy, value-added, timeliness, completeness, amount of data Representational DQ Interpretability, ease of understanding, concise representation, consistent representation
18.5.1.2 Criteria of Flexibility Heterogeneity and dynamism are cornerstone characteristics of autonomous cooperating logistics processes. Consequently, the second category of criteria refers to the ability of a data integration mechanism to, on the one hand react flexibly to changes in the surrounding system environment, and on the other be applicable to disparate types of data source.
Data Source Scalability Data Source Scalability refers to the ability of a data integration mechanism to facilitate the growth of the surrounding system. Scalability in this sense refers to the data administration difficulties in creating and maintaining large systems [25]. The central issue addressed revolves around the administrative effort required to add new data sources to an integrated system. This criterion directly impacts on the scalability of autonomous cooperating logistics systems. For example, if a restriction is put on the scalability of the underlying data integration mechanisms, this limit may also apply to the introduction of new intelligent logistics objects.
Data Volume Scalability In contrast to data source scalability, this dimension of scalability describes the ability of a mechanism to handle increases in data volume with regards to its run-time performance.
Data Source Agnosticism Data source agnosticism describes the capability of a given data integration mechanism to be applied to different types of data source and interface (cf. Table 18.1), if necessary simultaneously. Due to the expected heterogeneity, hierarchy and degree of distribution of IT systems and data sources in logistics systems exhibiting a high degree of autonomy, this capability impact directly on a mechanism’s ability support such systems. For example, a specialised mechanism which is capable
262
K.A. Hribernik et al.
of adequately integrating enterprise systems may be inadequate for the integration of sensor data. Such a mechanism alone would not do the requirements toward heterogeneity justice.
Adaptability Adaptability describes several dimensions of a given data integration mechanism in its relation to autonomous cooperating logistics processes. Foremost, it refers to the ability of the mechanism to adapt changes in the data sources in the IT logistics landscape. This can mean adding or removing a data source to or from the pool of integration targets. For example, an intelligent logistics object such as a parcel is introduced into the IT landscape of a logistics provider it has hitherto not been involved with. It may need, for example, to be able to access that provider’s IT systems to identify a suitable means of transportation. Using a closely-coupled approach to data integration, adding such a new data source would mean the modification of the federated data scheme. This would result in considerable effort. Using a loosely-coupled approach, the integration could take place immediately, but the responsibility for interpreting the data and solving heterogeneity conflicts would be placed with the intelligent logistics object.
18.6 Results: Limitations on Autonomous Cooperating Logistics Processes This section presents an evaluation of the data integration mechanisms discussed previously against the criteria defined in the previous section. An overview of the evaluation is presented in Table 18.4, which is elaborated in the following. Although approaches to realising near real-time data warehouses exist, the decision-making process in traditional data warehouse environments is often delayed because data cannot be propagated from the source system to the data warehouse in time [5]. This is especially disadvantageous to autonomous cooperating logistics processes – the degree of autonomous control increases with the dynamism of its decision system [40]. Since operational data stores are very similar to data warehouses, the evaluation of ODS is very similar to the results of the data warehouses. Their data timeliness is slightly superior, because the interval between requesting and storing new data sets from the data sources is typically much smaller. But this also leads to a lack of data volume scalability, because only the most current data sets are available. Older data records, as needed for object tracking for example, aren’t available. Traditional, tightly-coupled FDBS are robust, response and data timely. Due to the strict definition of a federated data scheme over all local schemata, a high level of data quality may be guaranteed. However, data source scalability is not an advantage of these FDBS – each time a new data source of a new type is added, the federated
18
Impacts of Data Integration Approaches
263
Semantic Mediator
Object-oriented Databases
Peer Database Systems
Enterprise Service Bus
Message-oriented Middleware
Service -oriented Architecture
Timeliness
Federated Database Management System
Criterion
Reliability
Operational Data Store
Category
Data Warehouse
Table 18.4 Evaluation of data integration mechanisms
Reactivity Robustness Quality of Data Flexibility
Data Volume Scalability Data Source Scalability Data Source Agnosticism Adaptability weakly,
slightly,
averagely,
largely and
strongly fulfilled
data schema needs to be modified. It also needs to be changed each time local data schemata are altered. SOA-based integration approaches are highly flexible towards data source scalability, agnosticism towards data sources and changes made to local data sources. Services are implemented at the local systems and their specifications published in repositories. Services interfacing new or modified data sources can thus be quickly implemented and published, ready for immediate integration. A system integrated using SOA is highly robust against the failure of individual services. The drawback of this approach is that the integration effort is placed upon the querying application. Furthermore, the large amount of data overhead generated for each service call and response using, for example SOAP RPC [13], make SOA less scalable with respect to data volume. The Enterprise Service Bus provides good data timeliness because the data can be queried directly from the data sources instead of querying local images like in data warehouses. With the use of SOA the ESB can be referred to be robust. There are disadvantages in ESB with response timeliness (messages can decay in messages queues) and data volume scalability (high data overhead and transformations in communication between components).
264
K.A. Hribernik et al.
The main advantages of object oriented (multi-)databases are the data timeliness (ad-hoc querying of data) and quality of data (data transformation and cleaning mechanism can be applied). Similar to FDBS, the object oriented databases constitute a global schema which narrows down their adaptability and data source scalability. Semantic Mediators have advantages in timeliness, quality of data, data source scalability and adaptability. Like many of the other approaches, the semantic mediator queries the data sources at query time, thus the timeliness is very good. The amount of data sources integrated using a semantic mediator can be theoretically infinite. This is because the mediator queries only data sources which may supply data according to the initial user-query, thus there is only little overhead with unaffected data sources. Due to the loose coupling of data sources in a semantic mediator, the adaptability is also very good. The weak spots of the mediator are the reactivity and data volume scalability. The mediator queries the data sources at query time, whereat the responses may take a while. In addition this implies that most of the integration tasks, like data transformation or duplicate recognition for example, have to be done at query time also. Hence, huge amounts of data sets increase response time.
18.7 Summary and Conclusions This contribution has endeavoured to systematically analyse the limitations imposed by different data integration approaches on autonomy in autonomous cooperating logistics processes, and to establish a scheme of categorization for identifying adequate data integration approaches for different degrees of autonomous cooperation. To achieve this, first possible classification of data integration approaches was presented. Then, the major approaches which show potential for the application to the data integration in autonomous cooperating logistics processes were discussed. Subsequently, a catalogue of criteria for gauging the impact of specific data integration approaches on the degree of autonomy cooperating logistics processes exhibit was proposed. The criteria were grouped according to whether they relate to the reliability or flexibility of the data integration mechanism. The former encompass timeliness, reactivity, robustness and quality of data. The latter consist of data volume scalability, data source scalability, data source agnosticism and adaptability. Finally, the data integration mechanisms discussed previously were evaluated in the light of these criteria. The results of the evaluation do not disqualify any of the mechanisms from being used in information systems supporting autonomous cooperating logistics processes. They do, however, provide a guide to gauging what effects a data integration mechanism may have on the degree of autonomy in such processes and for what reasons. They also allow for an informed decision on a case-by-case basis as to which data integration mechanism is most suitable for which scenario of autonomous control.
18
Impacts of Data Integration Approaches
265
For example, in a scenario with few data sources and little fluctuation in the amount and type of intelligent logistics object involved, but high demands towards reactivity, robustness and quality of data, either a data warehouse or operational data store might be a sensible choice. However, it is also clear that this type of data integration mechanism will further limit the degree of autonomy that scenario will exhibit. On the other end of the spectrum, an information system for the support of a highly dynamic, heterogeneous and fluctuating IT landscape involving many different intelligent logistics objects will necessitate a different data integration approach. Here, an approach which strongly supports data source scalability, agnosticism and adaptability is preferable. A number of approaches such as SOA-based integration mechanisms and semantic mediation fulfil these requirements. Both also strongly support timeliness and fulfil other criteria well. Consequently, the use of these data integration approaches is potentially less limiting on the degree of autonomy the autonomous cooperating logistics system may exhibit.
References 1. Aberle G (2003) Transportwirtschaft: einzelwirtschaftliche und gesamtwirtschaftliche Grundlagen, 4th edn. Oldenbourg, Munich 2. Bancilhon F (1988) Object-oriented database systems. In: Proceedings of the seventh acm sigact-sigmod-sigart symposium on principles of database systems. Association for Computing Machinery, Austin 3. Baragoin C, Marini M, Morgan C, Mueller O, Perkins A, Yim K (2001). Building the operational data store on DB2 UDB. IBM, San Jose, CA 4. Böse F, Windt K (2007) Catalogue of criteria for autonomous control in logistics. In Hülsmann M, Windt K (eds) Understanding autonomous cooperation & control in logistics – the impact on management, information and communication and material flow. Springer, Berlin, pp 57–72 5. Bruckner R, List B, Schiefer J (2002) Striving towards near real-time data integration for data warehouses. In: Goos G, Hartmanis J, van Leuwen J (eds) Lecture notes in computer science, vol. 2454/2002. Springer, pp 173–182, doi:10.1007/3–540–46145–0_31 6. Calvanese D, De Giacomo G (2005) Data integration: a logic-based perspective 7. Curry E (2004) Message-oriented middleware. In Mahmoud QH (ed) Middleware for communications. Wiley, pp 1–28 8. Degenring A (2005) Enterprise service bus. Java Spektrum 16–18 9. Dogac A, Dengi C, Özsu MT (1996) Building interoperable databases on distributed object management platforms. Commun ACM 10. Ehnert I, Arndt L, Mueller-Christ G (2006) A sustainable management framework for dilemma and boundaries in autonomous cooperating transport logistics processes. Int J Environ Sustain Dev 5:355–371. doi:10.1504/IJESD.2006.011555 11. Endrei M, Ang J, Arsanjani A, Chua S, Comte P, Krogdahl P, et al. (2004) Patterns: serviceoriented architecture and web services. IBM 12. EPCglobal Inc. (2007) EPC information services (EPCIS) Version 1.0.1 Specification. http://www.gs1.org/sites/default/files/docs/epcis/epcis_1_0_1-standard-20070921.pdf. Accessed 23 Dec 2010 13. Gudgin M, Hadley M, Mendelsohn N, Moreau J-J, Nielsen HF (2003) SOAP Version 1.2 Part 1: Messaging Framework, W3C Recommendation, World Wide Web Consortium. http://www.w3.org/TR/soap12-part1/. Accessed 06 Sept 2010
266
K.A. Hribernik et al.
14. Gupta A, Ludäscher B, Martone ME (2003) Registering scientific information sources for semantic mediaton. Lecture Notes Comput Sci 182–198 15. Hannus M (1996) Islands of automation in construction. In: Turk Žiga (ed) Construction on the information highway. CIB Publication, University of Ljubljana 16. Hans C, Hribernik KA, Thoben K-D (2008) An approach for the integration of data within complex logistics systems. In: Haasis HD, Kreowski H-J, Scholz-Reiter B (eds) Dynamics in logistics. First international conference LDIC 2007 proceedings. Springer, Heidelberg 17. Heimbinger D, McLeod D (1985) A federated architecture for information management. ACM Trans Inf Syst 253–278 18. Hribernik K, Hans C, Kramer C, Thoben K-D (2010) Data integration for an internet of things supporting autonomous cooperating logistics processes. In: Uckelmann D, Harris M, Michahelles F (eds) Architecting the internet of things. Springer, Berlin 19. Inmon WH (2000) Information architecture for the 90’s: Legacy Systems. ODS and DW 20. Inmon W (1996) The data warehouse and data mining. Commun ACM 49–50 21. Menge F (2007) Enterprise service bus. In: Free and open source software conference 22. Natis YV (2003) Service-oriented architecture scenario. Gartner, Stamford 23. Ortiz SJ (2007) Getting on board the enterprise service bus. Computer 15–17 24. Pitoura E, Bukhres O, Elmagarmid A (1995) Object orientation in multidatabase systems. ACM Comput Surv (CSUR) 25. Rosenthal A, Seligman L (2001) Scalability issues in data integration. AFCEA Federal Database Colloquium, San Diego 26. Roth A, Naumann F (2006) Benefit and cost of query answering in PDMS. In: Moro G, Bergamaschi S, Joseph S, Mori J-H, Ouksel AM (eds) Databases, information systems, and peer-to-peer computing, international workshops. Springer, pp 50–61 27. Scholz-Reiter B, Windt K, Freitag M (2004) Autonomous logistic processes: new demands and first approaches. In: Monostori L (ed) Proceedings of 37th CIRP international seminar on manufacturing systems. Hungarian Academy of Science, Budapest 28. Scholz-Reiter B, Kolditz J, Hildebrandt T (2007) Specifying adaptive business processes within the production logistics domain – a new modelling concept and its challenges. In Hülsmann M, Windt K (eds) Understanding autonomous cooperation & control in logistics – the impact on management, information and communication and material flow. Springer, Berlin, pp 275–201 29. Sheth AP, Larson JA (1990) Federated database systems for managing distributed, heterogeneous, and autonomous databases. ACM Comput Surv (CSUR) – Special issue on heterogeneous databases 30. Soon TJ, Shin-ichii I (2007) EPCIS and its applications. Synth J 109–124. Information Technology Standards Committee. IDA Singaport & SPRING Singapore. http://www.itsc.org.sg/ pdf/synthesis07/Five_EPCIS.pdf. Accessed 06 Sept 2010 31. Strong DM, Lee YW, Wang RY (1997) Data quality in context. Commun ACM 103–110 32. Tatarinov I, Halevy A (2004) Efficient query reformulation in peer data management systems. In: SIGMOD ’04 proceedings of the 2004 ACM SIGMOD international conference on management of data 33. The PROMISE Consortium (2008) PROMISE architecture series volume 1: architecture overview. Promise Innovation Oy, Finland 34. Timm IJ (2006) Strategic management of autonomous software systems: overview article. Technical Report 35. University of Bremen, Centre for Computing and Communication Technologies (TZI) 35. United Nations Economic Commission for Europe (UNECE) (2010) United nations directories for electronic data interchange for administration, commerce and transport. http://www.unece.org/trade/untdid/welcome.htm. Accessed 23 Dec 2010 36. Wache H (2002) Semantische mediation für heterogene informationsquellen. Akademische Verlagsgesellschaft Aka GmbH, Berlin 37. Widom J (1995) Research problems in data warehousing. In: Proceedings of the fourth international conference on information and knowledge management. ACM, Baltimore, MD, pp 25–30
18
Impacts of Data Integration Approaches
267
38. Wiederhold G (1994) Interoperation, mediation, and ontologies 39. Wiederhold G (1992) Mediators in the architecture of future information systems. Computer 38–49 40. Windt K, Böse F, Phillipp T (2005) Criteria and application of autonomous cooperating logistics processes. In: Proceeding of the 3rd international conference on manufacturing research – advances in manufactoring technology and management, Cranfield 41. Windt K, Philipp T, Böse F (2008) Complexity cube for the characterization of complex production systems. Int J Comp Integr Manuf 21:195–200. doi:10.1080/09511920701607725 42. Windt K, Böse F, Philipp T (2008) Autonomy in production logistics – identification, characterisation and application. Int J Robot CIM 24(4):572–578 43. Xu L, Embley DW (2004) Combining the best of global-as-view and local-as-view for data integration. In: Information systems technologies and its applications – ISTA, pp 123–136 44. Ziegler P, Dittrich K (2004) Three decades of data intecration – all problems solved? In: IFIP international federation for information processing, vol 156/2004, pp 3–12. doi:10.1007/978– 1–4020–8157–6_1
Part IV
Practical Contributions and Limitations
Chapter 19
Considerations of Practical Contributions and Limitations of Autonomous Cooperation and Control in Logistics Michael Hülsmann, Philip Cordes, and Anne Schwientek
The previous chapters of this book showed, in what way the idea of autonomous cooperation and control of logistics processes can contribute to the efficiency and robustness of logistics systems. This also includes critical discussions of limitations for the achievement of logistics goals. First, the question has been addressed how the organization of logistics systems is affected by a shift to a higher degree of autonomous cooperation and control and which other effects, in turn, emanate from it. Second, methodologies have been investigated that enable autonomous cooperating and controlled logistics processes. Third, technical contributions and limitations for a successful application of autonomous cooperation and control have been analyzed. However, real applications, in which this idea is realized to its full extent in the logistics practice, are rare, and so are the insights about practical contributions and limitations. The intelligent container is one example for a technology whose features base upon the idea of an autonomous logistics object that is able to interact with other objects and to decide independently from a central management. It provides realtime information about the quality of transported perishable goods and quality alterations over time [5–7]. Technologically, the container could be able to render main decisions, formerly made by a central management, such as destination and route changes due to shortened shelf life of the transported goods [4]. From a theoretical point of view, this idea is based on a wide literature foundation predicting advantages and disadvantages of suchlike self-organized logistics systems (e.g. [2,3,8–10]). However, since logistics companies do not use the intelligent container yet in their common logistics processes, the question of its practical feasibility and usability is not fully answered up to now. Hence, the validity of the primary findings regarding the container’s potential benefits and costs has to be empirically investigated in order to deduce its practical contributions and limitations.
M. Hülsmann (B), P. Cordes, and A. Schwientek Systems Management, International Logistics, School of Engineering and Science, Jacobs University Bremen, Bremen, Germany e-mail:
[email protected],
[email protected],
[email protected] M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_19,
271
272
M. Hülsmann et al.
This example indicates that the transformation from theory to technology necessary for the deduction of well founded recommendations (on this see [1]) regarding the implementation of autonomous cooperation and control technologies and methodologies might not be completed up to now. A theoretical framework for a shift from centralized planning of logistics processes to decentralized control through autonomous logistics objects has been developed, as well as the necessary methodologies and technologies. However, the question, whether or not the developed technologies and methodologies are realizable in the logistics practice, and whether there is a general pattern observable that proves the idea of autonomous cooperation and control to be feasible and usable in the ‘real world’ is still on the research agenda. Therefore, the central research question of this chapter reads as follows: What are contributions and limitations of a feasible and reasonable application of the concept of autonomous cooperation and control in the logistics practice – in production as well as distribution logistics? Correspondingly, the central aim is to gain insights into the feasibility and the usability of the theoretically, methodologically and technologically developed concept of autonomous cooperation and control in real logistics applications. First, it is intended, to depict exemplary technologies and algorithms that enable the idea of autonomous cooperation and control to be realized in the logistics context. Second, different areas of application, in which autonomous cooperation and control-based methodologies and technologies have either been tested or which offer potentials for a successful implementation, will be described. These descriptions provide the basis for further analyses regarding the practicability of the idea of autonomous cooperation and control in logistics. Second, this chapter intends to reveal causal interrelationships between the application of autonomous cooperation and control-based methodologies (e.g. algorithms) respectively technologies and its practical feasibility and usability. These interrelations serve as the analytical basis on which practical contributions and limitations of a realization of autonomous cooperation and control in logistics processes can be revealed. Third, insights about specific contributions and limitations of realizing the idea of autonomous cooperation and control in the logistics practice shall be gained so as to be able to identify underlying patterns of a general kind. In order to achieve these aims, this chapter provides three sections, which show practical contributions and limitations of autonomous cooperation and control from three different perspectives. First, Till Becker and Katja Windt investigate in “A Comparative View on Existing Autonomous Control Approaches – Observations from a Simulation Study”algorithms that enable logistics processes to be autonomously controlled. A simulation study that is based on data, collected at a terminal of an automobile logistics company, shows that it is necessary to customize the application of autonomous cooperation and the used algorithms for each individual scenario. Hence, practical limitations of autonomous cooperation consist of the impossibility to generalize positive or negative effects of certain algorithms.
19
Considerations of Practical Contributions and Limitations
273
Second, Bernd Scholz-Reiter, Carmen Ruthenbeck, Michael Teucke, and Jantje Hoppert analyze “Limitations of Autonomous Control in Practical Applications – Report on Lessons Learned from Vehicle and Apparel Logistics”. An exemplary requirement for a successful implementation into the daily industrial routines is for instance the necessity that the available technological infrastructure suits the requirements given by the method used. A practical limitation is therewith that a company overlapping generalization of the requirements for methods or technologies that enable autonomous cooperation and control does not seem to be possible. Third, Marius Veigt, Farideh Ganji, Ernesto Morales Kluge and Bernd ScholzReiter investigate in the section “Autonomous Control in Production Planning and Control – How to Integrate Autonomous Control into Existing Production Planning and Control Structures” structural changes that occur when autonomous cooperation is implemented into an existing production system. The study shows that the availability of necessary information is one essential aspect for a successful execution of the process control in production logistics by an autonomous logistics object. Practical limitations result therewith from the high quantity and quality of sensors necessary for assuring that relevant information is provided.
References 1. Chmielewicz K (1979) Forschungskonzeptionen der wirtschaftswissenschaft, 2nd edn. Poeschel, Stuttgart 2. Hongler M, Gallay O, Hülsmann M, Cordes P, Colmorn R (2010) Centralized versus decentralized control – a solvable stylized model in transportation. Phys A 389(19):4162–4171 3. Hülsmann M, Illigen C, Korsmeier B, Cordes P (2010) Risks resulting from autonomous cooperation technologies in logistics. In: Li Y, Desheng W (eds) Proceedings of the 2010 IEEE international conference on advanced management science, pp 422–426 4. Hülsmann M, Lang W, Colmorn R, Jedermann R, Illigen C, Cordes P (2010) Ecological impacts of autonomously co-operating technologies in international supply networks – establishing a model for analyzing CO2 -effects of implementing intelligent containers in fruit logistics. In: Pawar KS, Lalwani CS (eds) Proceedings of the 15th international symposium on logistics: configuring next generation supply chains, pp 398–406 5. Jedermann R, Lang W (2008) The benefits of embedded intelligence – tasks and applications for ubiquitous computing in logistics. In: Floerkemeier C, Langheinrich M, Fleisch E, Mattern F, Sarma SE (eds) The internet of things – first international conference (IOT 2008) proceedings. Springer, Berlin, Heidelberg, New York, pp 105–122 6. Jedermann R, Lang W (2007) Semi-passive RFID and beyond: steps towards automated quality tracing in the food chain. Int J Radio Freq Ident Technol Appl 1(3):247–259 7. Jedermann R, Ruiz-Garcia L, Lang W (2009) Spatial temperature profiling by semi-passive RFID loggers for perishable food transportation. Comput Electron Agr 65(2):145–154 8. McKelvey B, Wycisk C, Hülsmann M (2009) Designing an electronic auction market for complex ‘smart parts’ logistics: options based on LeBaron’s computational stock market. Int J Prod Econ 120(2):476–494 9. Windt K, Hülsmann M (2007) Changing paradigms in logistics – understanding the shift from conventional control to autonomous cooperation and control. In: Hülsmann M, Windt K (eds) Understanding autonomous cooperation & control – the impact of autonomy on management, information, communication, and material flow. Springer, pp 1–16 10. Wycisk C, McKelvey B, Hülsmann M (2008) “Smart parts” supply networks as complex adaptive systems: analysis and implications. Int J Phys Distrib Logist Manag 38(2):108–125
Chapter 20
A Comparative View on Existing Autonomous Control Approaches: Observations from a Simulation Study Till Becker and Katja Windt
20.1 Introduction The concept of autonomous control offers additional possibilities to cope with fluctuating market conditions that have evolved over the recent years. Customers expect a variety of customized products and a fast product delivery, while a rising number of world-wide acting companies operate complex logistic networks around the globe [4]. Such an interlinked network of logistic processes that belong to numerous companies lead to an enormous increase in dynamics and complexity regarding the ability to coordinate, control, and monitor these processes. Traditional production planning is likely to be disadvantageous in a situation that is characterized by high complexity and dynamic behavior. The reason can be found in the fact that an increase in complexity (e.g. number of links in a network, number of choices regarding variants) enables an exponential rise in possible combinations within a production program. Dynamic behavior affects production planning because changes in the environment occur while the “optimal” production program is still being calculated. This constant change of parameters, which are at the same time input parameters of the planning algorithm, renders the outcome of the planning process useless. Therefore, a different approach to complex and dynamic production planning problems is needed to overcome these two major obstacles. Autonomous control in logistics is one possible answer to the problems of complexity and dynamics. It is “characterized by the ability of logistics objects to process information, to render and to execute decisions on their own” [12]. Consequently, there is a spatiotemporal transfer of production planning decisions. Firstly, decision making now takes place within the logistic process itself, e.g. on the shop floor, instead of being carried out by a central department or a central IT system. Secondly, decisions are taken continuously during the process, ideally instantaneously before its execution in order to retain as much flexibility as possible [13]. T. Becker (B) and K. Windt Global Production Logistics, International Logistics, School of Engineering and Science, Jacobs University Bremen, Bremen, Germany e-mail:
[email protected] M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_20,
275
276
T. Becker and K. Windt
Classic production program planning realizes decision making prior to the execution of the complete production program, therefore having a clear temporal delay between decision making and decision execution. Previous publications have already addressed the question whether there is a benefit for the application of autonomous control in producing companies [14]. Additionally, the mapping of specific autonomous control algorithms to specific areas of application has been investigated [14]. The present work will extent the scope of the previously published studies. It will shed light on the boundaries that companies are expected to face when applying autonomous control in their logistic processes. The questions that are going to be addressed are as follows: Which autonomous control methods can be beneficial in different settings of
production processes? What are the limitations of the investigated autonomous control methods regard-
ing logistic targets achievement? The following sections of this paper are guided by the two questions. Autonomous control methods including the algorithms behind autonomous control, are analyzed more thoroughly in Sect. 20.2. Section 20.3 contains a description of the simulation model that has been used to study the available autonomous control methods and their logistic performance, as well as a summary of the eight different control methods that have been implemented. In the Sect. 20.4, the experimental data is evaluated and interpreted in the context of the two research questions. Finally, the answers to the research questions are summarized in Sect. 20.5.
20.2 Autonomous Control Methods The introduction of autonomous control in a production logistics company has organizational, technological, and process-related demands [4]. Organizational demands cover the need of autonomous logistic process description, the availability of local information, and more. Technical demands can be information processing ability of the autonomous actors, communication technology within the processes, and similar technology related requirements. Process-related demands include the actual control algorithms that determine how decision making is carried out. These algorithms are called autonomous control methods. They are defined according to the definition of autonomous control as “generic algorithms that describe how logistics objects render and execute decision by their own” [15]. Autonomous control methods can have any degree of complexity regarding their decision making. A simple instance of an autonomous control method could allow semi-finished parts in a job-shop scenario to select the next step in their production by choosing the machine having the lowest number of items waiting in its queue [5]. A more sophisticated method is inspired by ants’ foraging behavior. A preferred path through the production is indicated by virtual pheromones that are emitted by successfully processed parts [1].
20
A Comparative View on Existing Autonomous Control Approaches
277
Beside the work this paper builds upon [14], there have been other analyses of autonomous control methods. Scholz-Reiter et al. (2009) conducted a simulation study as well, but compared only three different autonomous control methods [8]. Furthermore, they focused on the logistic target achievement in distinct scenarios with varying process complexity and degree of autonomous control. Peng and Mcfarlane (2004) defined three different strategies for agent based manufacturing management [3]. Their simulation was based on a general flow shop model. Their results indicate that going for the shortest waiting queue is a dominant solution compared to breakdown avoidance and proportional availability strategies. Other simulations [6, 7] were utilized to prove the feasibility of newly developed autonomous control methods without a comparative aspect. A recent publication of Scholz-Reiter et al. (2010) investigates autonomous control in comparison to common scheduling heuristics [9]. In contrast to the approach presented here, they utilized an exemplary production process model. A previous study of autonomous control methods has shown that the existing methods can be sorted in clusters which have common characteristics [15]. This has been so far the investigation that incorporated the largest number of methods in the area of autonomous control in logistics. Fourteen available methods have been categorized and evaluated in a simulation study. However, a simplified simulation model representing a job shop production scenario has been applied there. As the benefits and drawbacks of autonomous control methods in production logistic processes are brought into focus in this work, a more realistic simulation model has been used here. Furthermore, this model is derived from a real business case in production logistics. This yields to a higher confidence of the simulation results and their interpretation regarding the application of autonomous control methods in industry.
20.3 Simulation Model 20.3.1 Model Description The simulation model designed for this study is based on the automobile terminal at the Hamburg port in Germany. The terminal is operated by the BLG LOGISTICS GROUP. The terminal is a transshipment point as well as a production site. In the highly standardized automotive industry car manufactures outsource non-standard modifications to contractors. The automobile terminal offers storage capacity for 12,000 cars on 324,000 square meters. Beside storage areas the terminal has technical facilities which allow the operator to conduct repairs, painting, washing, and technical modifications. The terminal receives cars via vessel, train, or truck. The distribution is usually done by truck. In addition to the availability of the terminal data, the scenario has been selected for the simulation study due to the comparatively simple applicability of autonomous control. The processes on the terminal have a clear structure and all elements and actors are easily identifiable. The process-inherent flexibility regarding
278
T. Becker and K. Windt
the sequence of the production steps enables autonomous, ad-hoc decisions. This process represents a class of multistage production processes. The lifecycle of a car on the terminal starts with its arrival via vessel, train, or truck. All incoming cars are taken to the Incoming Deliveries Area. They are registered electronically according to their chassis number. The workers do a quick check and a damage survey if required. The car ID is linked to the orders which are already available in the IT system. Cars can have no orders at all and are supposed to be directly shipped to their final destination, which is usually a car dealer. Otherwise, there is a parking order and/or one or more treatment orders. When the receiving car dealer needs a car that is designated to him, he orders the car from the terminal. A retrieval order is created and the car is retrieved from storage, sent through the treatment if applicable, and placed on the Outgoing Deliveries Area. The necessary documents are created and the car leaves the terminal. A schematic layout of the terminal is displayed in Fig. 20.1. In order to answer the two research questions, it is necessary to model the relevant aspects of the terminal’s logistic processes as they take place in reality. For purposes of the simulation study, the control mechanism as it is used on the terminal is replaced by the selected autonomous control methods. Data from key figures is selected and evaluated according to the initial questions. The complete production data of the year 2008 serves as input data for the simulation study. The dataset includes 151,934 car records. Each record is composed of the arrival date, information on parking order including duration of parking, and a list of treatment orders. The path of the virtual cars in the simulation model is depicted in Fig. 20.2. All cars are registered at their arrival date at the Incoming Delivery Area. The process
Parking Area 5
Parking Area 4
Parking Area 2 Parking Area 1
Outgoing Deliveries
Incoming Deliveries
Hall 6
Gas
Washing
TPR Parking Area 3
Parking Area 9
Parking Area 6
Hall 8
Waiting Treatment
Parking areas
Parking Area 7
Hall 1
Parking Area 8 Diesel
TPR = Transport Protection Removal
Buildings containing treatment facilities
Fig. 20.1 Schematic representation of the automobile terminal
20
A Comparative View on Existing Autonomous Control Approaches
Incoming cars
Parking Area 1
Treatment 1
Parking Area 2
Treatment 2
…
Buffer
…
…
…
…
…
Parking Area 9
Treatment 8
279
Outgoing cars
Fig. 20.2 Possible paths of a car in the simulation model
of transportation is not modeled explicitly as it has no relevance for the processing of the orders. Each car can either immediately be taken to the Outgoing Delivery Area, be transferred to one of the parking areas, or be transferred to the Treatment Waiting Area which serves as a buffer for all treatment stations. The choice depends on the existing orders for each individual car. Cars that are retrieved from a parking area either go to the Treatment Waiting Area if a treatment order is available or they go directly to the Outgoing Delivery Area and thus leave the simulation. Cars never return to the parking areas as treatment only takes place immediately before they are shipped. All cars that have at least one treatment order wait on the Treatment Waiting Area (buffer). After one single treatment has been finished cars enter again in the buffer if there are remaining unfinished treatment orders. Otherwise the car is transferred to the Outgoing Delivery Area. The time for transferring a car from one point to the terminal to another is taken from a distance table. The table contains average values from all sources to each applicable destination. The treatment times are determined by randomly drawing a value from a table containing all treatment times for each station taken from the original dataset. Consequently, the treatment times in the simulation follow the same distribution as the times from the 2008 dataset.
20.3.2 Autonomous Control Methods Applied The comparison of the different autonomous control methods is realized by embedding the control mechanism of each single method into the simulation model. For the selection of appropriate autonomous control methods, a thorough literature study
280
T. Becker and K. Windt
has been performed. Control methods in the areas of autonomous control, selfcontrol, multi-agent-based methods, smart systems, or decentralized control have been taken into consideration. There were two main criteria for the selection of a method: firstly, the methods need to be clearly documented so that it is possible to implement them in the simulation model. If the concept of a method is only described as a sketch, it cannot be guaranteed that it is implemented the way it is supposed to work. Secondly, the methods need to be applicable to the present simulation model. A too narrow scope of application (e.g. a restriction to transportation or assembly processes) would make the control methods useless for this scenario. Autonomous control methods have been collected within the collaborative research center “Autonomous Cooperating Logistic Processes” (CRC637, see Acknowledgement section) as well as from external sources. Table 20.1 gives an overview of the eight selected methods. The parameter “depth” which is given in the short description of the autonomous control method describes how many steps in advance are considered by the specific method [16]. For example a depth of two when using queue length as decision criterion indicates that the car would sum up the queue lengths of two following stations, thus creating a decision tree. Each decision alternative then consists of two consecutive treatments, valued by their summed queue lengths.
Table 20.1 Autonomous control methods applied in the simulation study. Parameter depth indicates how many following processing steps are taken into consideration for the decision (full means that all remaining steps were evaluated) [14] Name Source Short description Standard (std) Automobile Fixed assignment of parking area terminal Fixed sequence of treatment steps DLRP (dlrp) [11] Choose path with lowest estimated travel time to parking/average travel C waiting C processing time; depth: full Ant Pheromone [6] After each car departs from a treatment station it leaves a fixed (ant) amount of pheromones Pheromones add up with the already existing amounts at the station; over time the pheromones slowly evaporate Stations with higher pheromone levels are stations with better throughput and are therefore preferred by the cars; depth: 4 Holonic (holonic) [10] Choose path with lowest estimated travel time to parking/average travel C waiting C processing time; depth: 2 Minimum buffer CRC637 Choose parking area/treatment station with lowest occupation (minbuf) (relative buffer level); depth: 2 Queue Length CRC637 Choose path with lowest queue length (# of items); depth: 1 Estimation (qle) Random (random) CRC637 Random assignment of parking area Random sequence of treatment steps Simple rule [5] Choose path with lowest estimated travel time to parking/waiting based (sr2) and processing time; depth: 4
20
A Comparative View on Existing Autonomous Control Approaches
Table 20.2 Priority list parking areas Priority With treatment 1 Parking area 4 2 Parking area 5 3 Parking area 6 4 Parking area 7 5 Parking area 8 6 Parking area 9 7 Parking area 2 8 Parking area 3 9 Parking area 1
281
Without treatment Parking area 1 Parking area 3 Parking area 2 Parking area 4 Parking area 5 Parking area 6 Parking area 7 Parking area 8 Parking area 9
Table 20.3 Priority list treatment stations Rank Treatment station 1 Gas or diesel 2 Transport protection removal 3 Washing 4 Hall 6 5 Hall 8 6 Hall 1
The standard (std) method describes a set of rules which is based on the procedures that have been applied at the automobile terminal at the time when processes and data have been recorded. The parking areas are filled in a specific order. The order depends on the existence of one or more treatment orders for a car. Cars without treatment orders are assigned to different parking areas. All treatment orders, if available, are executed in a fixed sequence. Tables 20.2 and 20.3 show the priority lists applied in the simulation for the standard method. Another special method is the random method. This method is not an autonomous control method in the sense of a specific strategy. It simply chooses the following step (parking area or treatment station) by random selection. The selection follows a uniform distribution over the number of alternatives available, not considering the capacity. The two methods, standard and random, have been implemented in the simulation in order to let the autonomous control methods compete against them.
20.3.3 Data Evaluation Scheme A performance evaluation can only be done if a key figure measurement is available. The four targets of production logistics (short throughput time, high due date reliability, high utilization, and low work in process) can serve as a measurement system in many logistic applications [2]. In the simulation model throughput time has been measured in separate portions. They consist of travelling time, parking time, and treatment time. As the real due dates could not be derived from the available dataset,
282
T. Becker and K. Windt
the due date for each car has been set to 4 working days after its retrieval date. The retrieval date is recorded in the original dataset. It is the point in time when the customer (usually the retailer) demands the car. At that point a car is transferred initially to the treatment area, either from the parking area or directly from the incoming area. Both throughput time and due date reliability are based on actual working shifts, excluding nights, weekends, and holidays. Utilization as well as inventory levels of parking areas and treatment stations have been collected on a daily basis. Work in process as a key figure has been left out because due to the static input data the average work in process over the whole year remains unchanged, regardless of the control method. Although it would be difficult to draw conclusions about exact performance values of the automobile terminal from the simulation study, the values for throughput and travel times have been multiplied with a hidden factor to protect the company’s business data. The due date reliability as it is calculated here is a realistic assumption but the individual due dates are part of non-public contracts of the company. The specific due date reliability figures in this study do not necessarily need to match the companies true due date reliability. However, the approach used here allows for a judgment of due date reliability relations between the tested methods.
20.3.4 Simulation Model Parameters The terminal model and the given input data for the year 2008 represent a fixed environment in which the different autonomous control methods have been investigated. But one important point of autonomous control and self-organization is the fact that the methods are meant to adapt to different situations. On the one hand, this happens over time depending on the varying arrival of cars over the year. On the other hand, one of the initial questions in this work aims at getting a better understanding under which circumstances a specific autonomous control methods performs better or worse. Therefore, additional simulation runs with a variation of selected model parameters have been carried out. The scenario as it has been described up to now is called the basic scenario. This includes the presented model and the original 2008 production data. The second scenario is called “full flexibility” [17]. In this setting the restrictions regarding the sequence of treatment orders have been suspended. Consequently, the average number of decision alternatives for each car increases, resulting in a higher flexibility of the products. Although this does not represent the actual situation in this real world application of autonomous control methods, it gives a hint how the performance of the methods changes under varying flexibility. A third scenario investigates the methods’ performance depending on the workload. In this “increased load” scenario the treatment times have been increased by 10%. Consequently, the capacity of the whole system is reduced. Finally, a fourth scenario combines both modifications and includes higher flexibility and increased workload.
20
A Comparative View on Existing Autonomous Control Approaches
283
For each combination of the four scenarios with each of the control methods, five independent simulation runs have been carried out in order to ensure the significance of the results. Statistical tests have been applied to verify the significance of the deviations between the figures presented in Sect. 20.4. The error values have not been included in the histograms due to the high spread of the individual values. This spread, however, is not caused by an arbitrariness of the model. It is a result of the vast variation in the input data between the individual cars regarding the number of orders, the individual treatment times, and the fluctuating workload over the year.
20.4 Simulation Results 20.4.1 Utilization The simulation results will be presented according to the previously mentioned production logistics targets. One of the targets is to keep utilization of the resources at a high level. Resources in the simulation model are the parking areas, which correspond to storage facilities on a more generic level. The automobile service provider seeks to fully utilize its parking areas, because the customers pay fees for the storage of their cars on a daily basis. The treatment stations can be seen as machines or assembly stations in terms of adding value to a product. Only the parking area utilization is presented here, because due to the fixed structure of the input data (fixed number of cars, fixed assignment of treatment orders) there was no variation in the treatment station utilization figures. Furthermore, the scenario modifications did not affect parking area utilization, so that only results from the standard scenario are presented here. Figure 20.3 shows a graph indicating the average parking area utilization in the simulation model over the whole year. The methods ant and qle have not been included in the utilization evaluation as these methods are not applicable to the parking area selection decision. The virtual pheromone approach used by the ant method is depending on the evaporation of the pheromones over time. However, the time a car spends on the parking area is externally determined and therefore the evaporation would not represent waiting time as the parking time biases the decision variable pheromone level. The qle method bases its decisions on the lengths if waiting queues occur. The lack of waiting queues at the parking areas renders the decision method useless at this point. In the simulation, the parking decisions for these two methods have been substituted by the minbuf method. There are different characteristic patterns of parking area utilization that can be observed in Fig. 20.3. The dlrp and holonic method have almost the same distribution pattern of utilization, because they consider the same variables for their decision. Minbuf and sr2 distribute the cars equally between all storage areas, as these methods select the parking area with lowest utilization and therefore level the utilization over all areas. At first glance it seems surprising that random does not
284
T. Becker and K. Windt
Fig. 20.3 Average parking area utilization on the automobile terminal (adapted from Windt et al. 2010a)
distribute equally. The answer is that random makes an equally distributed choice between all parking areas, which can have different capacities. Consequently, high capacity parking areas are less utilized and vice versa. Std directs cars to the parking areas following the priority list given in Table 20.2.
20.4.2 Throughput Time: Travel The average times recorded for travelling to and from the parking area are displayed in Fig. 20.4. Again, the methods ant and qle which cannot make parking relevant decisions are left out. Values are shown for the basic scenario due to lack of significant differences between the different scenarios. It can be clearly seen that dlrp and holonic outperform the other approaches in terms of travel times. The belowaverage performance of minbuf, random, and sr2 are the trade-off for leveling the utilization of all parking areas, which also includes more distant parking sites.
20.4.3 Throughput Time: Treatment In the basic scenario (Fig. 20.5, grey columns) all methods except minbuf manage to outperform the simple std method regarding average throughput time. Curiously, even the random method delivers higher throughput time performance. If random is selected as benchmark, dlrp, holonic, and sr2 remain with increased performance. Interestingly, sr2 as a less sophisticated decision method in comparison to dlrp and holonic delivers the best result. This conveys the impression that in this scenario the average values used by dlrp and holonic do not match the actual situation exactly enough. The overall picture changes a bit when looking at the full
20
A Comparative View on Existing Autonomous Control Approaches
285
Fig. 20.4 Average travel times from the incoming area to parking and from parking to treatment (adapted from Windt et al. 2010a)
Fig. 20.5 Average treatment times of all cars that have at least on treatment order (basic scenario and full flexibility scenario)
flexibility scenario. The increased number of decision alternatives enables ant and qle to outperform the other methods. The information that is expressed by the virtual pheromone level in the ant method seems to reflect best the situation in the production process. Std remains unchanged due to the fixed allocation of order sequence, while random performs worst. Figure 20.6 depicts the average treatment times in the increased load and the combined scenario. Results are somehow similar for the increased load scenario. In the case of a combination of increased load and full flexibility, ant and dlrp are the only methods that can outperform the std approach. Again, the virtual pheromone
286
T. Becker and K. Windt
Fig. 20.6 Average treatment times of all cars that have at least on treatment order (increased load only and increased load & full flexibility scenario)
figure seems to offer a good indicator for throughput time decisions. Both dlrp and ant perform significantly better in a scenario with higher flexibility.
20.4.4 Due Date Reliability Due date reliability can be increased by any method in any scenario in comparison to the std approach. The difference among the methods decreases coming from the basic scenario to full flexibility (Fig. 20.7). Due date reliability drops in general under increased load (Fig. 20.8). Adding more flexibility lets the majority of the autonomous control methods perform better.
20.4.5 Opportunities and Limitations of Autonomous Control The simulation study delivered diverse results regarding the performance of autonomous control methods in a production logistics application scenario. First of all it has to be stated that there is no single autonomous control method that dominates all other methods in the achievement of all logistic targets in any scenario. A holistic view on the parking process (representing storage process in production logistics) includes the results from parking area utilization and travel times to and from parking. Dlrp and holonic manage to reduce travel times significantly and should be preferred among investigated methods for the purpose of storage. If the major goal is the balanced utilization of storage capacities, minbuf and sr2 perform best.
20
A Comparative View on Existing Autonomous Control Approaches
287
Fig. 20.7 Average due date reliability (basic scenario and full flexibility scenario)
Fig. 20.8 Average due date reliability (increased load only and increased load & full flexibility scenario)
The production (treatment) part needs more differentiation. The results vary depending on the available flexibility in the production process and the workload of the system. In general, dlrp has a good overall performance. However, in specific scenarios it can be significantly outperformed (e.g. by ant in full flexibility scenarios). Therefore, the results tend to suggest a distinct selection of the autonomous control algorithm, depending on various parameters, including logistic target preference, number of decision alternatives, and workload of the system. Limitations of autonomous control can be seen from two perspectives: firstly, the presented figures indicate that autonomous control methods can perform worse than static planning methods or even a random decision making. Consequently,
288
T. Becker and K. Windt
autonomous control does not inherently outperform other control approaches. A second perspective that might limit the application of autonomous control is the difficulty of selecting the appropriate autonomous control method. There is no single dominant method that can be applied in an arbitrary scenario. The introduction of autonomous control needs a careful judgment and customization. A possible solution can be the distinguished consideration of autonomous control methods for sub processes or different system states (e.g. high workload vs. low workload). Furthermore, a single autonomous control method could be parameterized to adapt itself to a changing environment. Finally it needs to be stated that this investigation used an inductive approach. The performance of the selected autonomous control methods has been observed in one specific scenario that represents a class of multistage production processes. Although the automobile terminal scenario has been carefully selected and implemented, there may be different results for the same autonomous control methods in distinct scenarios.
20.5 Conclusion This work has presented a simulation study based on a real world logistics scenario. Several autonomous control methods have been applied in the simulation and their performance regarding different logistic key figures has been analyzed. The results clearly point out that the application of autonomous control has to be customized for each individual scenario. The parameters of the logistic environment as well as logistic target preference determine which autonomous control method to use. It is also conceivable that different autonomous control methods are applied interchangeably in the same scenario, depending on surrounding conditions. Future research should aim at developing a clear structure for the classification of autonomous control methods regarding their parameters. In combination with additional simulation experiments in more diverse scenarios it will be possible to identify the high-performance methods with greater certainty.
References 1. Cicirello V, Smith S (2001) Ant colony control for autonomous decentralized shop floor routing. In: ISADS-2001, Fifth international symposium on autonomous decentralized systems 2. Nyhuis P, Wiendahl H (2008) Fundamentals of production logistics: theory, tools and applications. Springer, Berlin Heidelberg 3. Peng Y, Mcfarlane D (2004) Adaptive agent-based manufacturing control and its application to flow shop routing control. Prod Plann Contr 15(2):145–155 4. Scholz-Reiter B, Windt K, Freitag M (2004) Autonomous logistic processes: new demands and first approaches. In: Monostori, L (ed) Proceedings 37th CIRP international seminar on manufacturing systems. Hungarian Academy of Science, Budapest, Hungaria, pp 357–362
20
A Comparative View on Existing Autonomous Control Approaches
289
5. Scholz-Reiter B, Freitag M, de Beer C, Jagalski T (2006) The influence of production network’s complexity on the performance of autonomous control methods. In: Teti R (ed) Intelligent computation in manufacturing engineering, proceedings of the 5th CIRP international seminar on computation in manufacturing engineering (CIRP ICME ‘06). University of Naples, Naples, pp 317–320 6. Scholz-Reiter B, de Beer C, Freitag M, Jagalski T (2008a) Bio-inspired and Pheromone-based shop-floor control. Int J Comput Integrated Manuf 1(2):201–205 7. Scholz-Reiter B, Jagalski T, Bendul JC (2008b) Autonomous control of a shop floor based on bee’s foraging behavior. In: Proceedings of the first international conference on dynamics in logistics (LDIC), pp 415–423 8. Scholz-Reiter B, Görges M, Philipp T (2009) Autonomously controlled production systems – influence of autonomous control level on logistic performance. CIRP Ann Manuf Technol 58(1):395–398 9. Scholz-Reiter B, Rekersbrink J, Görges M (2010) Dynamic flexible flow shop problems – scheduling heuristics vs. autonomous control. CIRP Ann Manuf Technol 59(1):465–468 10. Van Brussel H, Wyns J, Valckenaers P, Bongaerts L, Peeters P (1998) Reference architecture for holonic manufacturing systems. PROSA Comput Ind 37(3):255–274 11. Wenning B-L, Rekersbrink H, Timm-Giel A, Görg C, Scholz-Reiter B (2007) Autonomous control by means of distributed routing. In: Hülsmann M, Windt K (eds) Understanding autonomous cooperation & control in logistics. Springer, Berlin Heidelberg, pp 325–335 12. Windt K, Böse F, Philipp T (2008) Autonomy in production logistics – identification, characterisation and application. Int J Robot CIM 24(4):572–578 13. Windt K, Jeken O (2009) Allocation flexibility – a new flexibility type as an enabler for autonomous control in production logistics. In: 42nd CIRP conference on manufacturing systems, Grenoble 14. Windt K, Becker T, Kolev I (2010a) A comparison of the logistics performance of autonomous control methods in production logistics. In: Sihn W, Kuhlang P (eds) Proceedings of the 43rd CIRP international conference on manufacturing systems: sustainable production and logistics in global networks. NW Verlag, Vienna, pp 576–583 15. Windt K, Becker T, Jeken O, Gelessus A (2010b) A classification pattern for autonomous control methods in logistics. Logist Res 2(2):109–120 16. Windt K, Becker T, Asenov D, Arbabzadah F (2010c) A generic implementation approach of autonomous control methods in production logistics. In: Proceedings of the 8th IEEE international conference on control & automation, Xiamen, China, 9–11 June 2010, pp 629–633 17. Windt K, Jeken O, Becker T (2010d) Selbststeuerung in der produktion – verbesserte logistikleistung durch ausschöpfung von flexibilitätspotentialen in fertigung und montage. ZWF Zeitschrift für wirtschaftlichen Fabrikbetrieb 105(5):439–443
Chapter 21
Limitations of Autonomous Control in Practical Applications: Report on Lessons Learned from Vehicle and Apparel Logistics Bernd Scholz-Reiter, Carmen Ruthenbeck, Michael Teucke, and Jantje Hoppert
21.1 Introduction Autonomous control is the research topic of the Collaborative Research Centre 637 (CRC 637). The CRC 637 defines autonomous control as processes of decentralized decision making in non-deterministic systems. An autonomous controlled system consists of interacting elements, which are characterized by their ability to process information, and to render and to execute decisions independently and on their own [4]. The application of autonomous control aims at increasing a system’s logistic performance and, in particular, its robustness, i.e. to improve a logistic system’s response to dynamic instability [9]. Until now, a completely successful implementation of autonomous control has not been achieved. However, different degrees of autonomous control have been achieved. The feasibility and usefulness of autonomous control in logistic systems is limited by different factors. In this paper we focus on methodical, technical, organizational and economic factors. These factors may increase the costs of implementing autonomous control in logistic systems, decrease its efficiency or effectiveness, or make it wholly unfeasible. For instance, in some scenarios, adequate decision methods for autonomous objects might be lacking. In other scenarios, these methods may exist, but the current technologies may not allow efficient and economic implementation. Limitations of autonomous control are considered non-permanent, but medium or long term in relation to time, being relevant for time frames of years or decades. This means they are more persistent than strictly temporal, short-term problems, which occur during introduction of or transition to autonomous control in an implementation project. Usually, such problems are encountered during implementation of autonomous control (as well as other process modifications) but can be quickly solved, within a couple of weeks or months. Lack of employees’ experience with new methods, or software bugs B. Scholz-Reiter (B), C. Ruthenbeck, M. Teucke, and J. Hoppert Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany e-mail:
[email protected] M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_21,
291
292
B. Scholz-Reiter et al.
may be pointed out as examples for such temporary problems. Not withstanding this, what we call limitations, are not, or at least not necessarily, permanent. Technical limitations, for instance, may be overcome by the introduction or maturation of a new technology, economic limitations may be overcome e.g. by changes in the availability of resources. In the following, the mentioned limitations are introduced. On the basis of two transfer projects, application specific, as well as common, limiting factors were encountered. These limitations, which influence the implementation of autonomous control, refer to either:
Methodical Technical Organizational, or Economic factors
The purpose of this article is to discuss limitations of autonomous control from a practical point of view. It focuses on experiences during the implementation process of autonomous control into industrial application. The two selected transfer projects took place in different application areas. One project dealt with vehicle logistics, a scenario already examined in previous theoretical work. The result of this project has been a wearable computing solution for vehicle allocation and tracking at an automobile terminal. The other project’s scenario is apparel logistics, which has been a new scenario and was not previously examined by the CRC 637. The aim of this project was to develop a method for autonomous disposition and allocation of articles to customer orders. Furthermore, this method was implemented into a prototypical solution including auto-identification technologies. Despite the different scenarios and objectives, some similar experiences were gained in these projects concerning limitations of autonomous control, which justify a common discussion in this article. On the one hand the discussion will be illustrate own specific aspects for each project and on the other hand the commonalities between the projects in a more general way. It has to be mentioned that experiences from only two projects cannot legitimately claim any general applicability. Nevertheless, we believe the gathered experiences are useful for other projects or applications of autonomous control in practical applications as well. The article is structured as follows. After this introduction, both application scenarios of the projects are introduced in the second chapter. The third chapter discusses different types of limitations of autonomous control encountered in those projects, both from a project specific as well as from a generalized perspective. The forth chapter tries to point out some general remarks on the limitations, discussed in the third chapter. The article closes with a conclusion and outlook in the fifth chapter.
21
Limitations of Autonomous Control in Practical Applications
293
21.2 Description of the Scenarios 21.2.1 Vehicle Logistics Scenario The project “Sensor Systems for Storage Management” belonged to the field of vehicle logistics. The scope of the project was the development of sensor systems for autonomous controlled storage management of vehicles at an automobile terminal. The basis for the transfer project was made by the fundamental research of the CRC 637. The fundamental research of decentralized decision making processes evaluated the implementation of rule-based autonomous control methods [31]. This evaluation demonstrated benefits for storage management of vehicles. The project aimed to examine the practical application of autonomous controlled storage management. Therefore, a hardware prototype with sensor systems and innovative information and communication technologies (ICT) was developed. In the following section the process of vehicle identification and storage allocation at the terminal will be explained. Figure 21.1 shows the scenario of vehicle logistics. Automobile terminals provide various services for vehicles in the range of transhipment, storage, and technical treatment. After unloading of a vehicle from ship, every vehicle is identified by a terminal employee using the vehicle
Storage Area
Unloading of Vehicles
Vehicle Identification
Wearable Computing System Technical Treatment Station
Fig. 21.1 Scenario of vehicle logistics
Disposition Area
294
B. Scholz-Reiter et al.
identification number (VIN). The VIN bijective defines every vehicle. For the following process steps a smart label a basis of the VIN is placed in the rear window on the offside of the vehicle. For every further vehicle movement this smart label identifies the vehicle. Afterward, a terminal employee moves the vehicle to the storage area. On demand of a car dealer, the vehicle proceeds to several technical treatment stations. After that, the vehicle is brought to the disposition area for transportation to the automobile dealer. The logistic objects are enabled to gather and process information, render and execute decision on their own [3]. The scope of investigation for autonomous control in this scenario is the storage management of vehicles [2]. On the one hand, the vehicles aim to achieve short transfer, on the other hand, storage areas try to maximize there storage occupancy rate. They offer the inquiring vehicles the total transfer time, which consists of the transfer time from the current vehicle location to the storage area, the parking time on the storage area as well as the future transfer time of the vehicle to the first technical treatment station. Depending on the stock level and the position of the storage areas in the automobile terminal, storage areas can offer a more or less convenient storage time and link to the next technical treatment station. Depending on the total transfer time, the vehicle chooses the best storage allocation. In order to implement the autonomous control method from the fundamental research on the automobile terminal, a prototypical wearable computing system was developed and the vehicles were equipped with smart labels with RFIDTransponders (Radio Frequency Identification). The wearable computing system takes different innovative Communication Technologies (ICT) components: A location module for determination of the position via Global Positioning
System (GPS). An identification module for the identification of objects via RFID. A communication module for data transfer via GPRS. A Thin-Film Transistor (TFT) display with touch-screen and audio hardware for
hand free user interaction. A proximity sensor for environmental information. An accumulator for securing the internal power supply.
The technical components are integrated in the professional clothing of the terminal employees. A schematic illustration is given in Fig. 21.2. The wearable computing system is the technical base for the integration of the autonomous control methods in the process of vehicle logistics in this scenario. For the process of vehicle movement, the wearable computing system identifies the vehicle using the RFID-transponder by the VIN. During the previously explained autonomous controlled decision making process the vehicle selects a storage area depending on the provided total transfer times. For this process the vehicle is represented by the wearable computing system. During the time when the RFIDtransponder is in the range of the system, the selected storage area is shown on the display of the wearable computing system. The terminal member moves the vehicle
21
Limitations of Autonomous Control in Practical Applications
295
GPS-Antenna Accumulator
ProximitySensor
GPRS/GSM-Module and -Antenna RFID-Reader & CPU-Module RFID-Antenna
Fig. 21.2 Sketch of the wearable computing system
to the assigned storage area. The wearable computing system observes the environment by sensor technique. Once the terminal member leaves the vehicle, the system saves the location by GPS. A special issue of wearable computing system is the high integration level of technology into the work process. Many of the previously developed wearable computing systems neglect the users’ special requirements, the supported process and the application area. Hence, they are rejected directly by the users or by employee organizations [15].
21.2.2 Apparel Logistics Scenario Besides vehicle logistics, apparel logistics was selected as an implementation scenario to test the applicability of autonomous control. The transfer project “Order Disposition in an Apparel Supply Chain” aimed to develop a method for allocation of article bundles to customer orders. In the scope of this project, disposition and control of customer orders and articles in an apparel supply chain have been investigated. Apparel logistics has to cope with limited sales periods and high product variability. Nevertheless, customers (or retailers) require short delivery times and high delivery flexibility. A special problem is caused by long delivery times due to long transport distances. Production has been outsourced to low wage countries (in particular China and South East Asia), while distribution and consumption take place in Europe and North America. Central research questions of the project were, whether autonomous controlled articles may be used in distribution processes, whether process improvements can be gained hereby and whether decentralization of necessary technical capabilities is technically feasible and economically sensible. The project’s objective was the development of a method for autonomous allocation of article bundles to customer orders during distribution processes. This allocation is based on consideration of compatibility of the ordered and the produced product variants, as well as on the spatial and timely availability in relation to the orders’ due dates. If necessary, the allocation should be dynamic, to react to short-term order modifications. Development of the method included selection of appropriate decision rules for article bundles and customer orders, as well as the implementation in a hardware and software based demonstration application.
296
B. Scholz-Reiter et al.
During the project a supply chain was inspected, which is similar to a reference chain described by Bruckner and Müller [5], with factories in China and Bangladesh. Central distribution centres are located in Germany, which serve retailers in Europe, Russia and the Near East. Articles are transported between factories and distribution centres via several hubs, with the main distance covered by container vessels (in urgent cases air transport may be used). Delivery times include up to ten weeks for raw material procurement and production, while transport times between factories and distribution centres are in between four to five weeks (values provided in [14]). Customer order specific article picking and dispatch is done after temporal storage of articles in the distribution centres. Standardized routing of all articles causes bottlenecks in the distribution centres, like short-term stock level peaks and work load peaks. Additionally, the number of different product variants is expected to rise with further individualization of customer requirements. For these reasons, the distribution network should be made more flexible, using new hub nodes for redirection of transports and delivering part of the articles directly to retailers without storage in distribution centres. Autonomous control is applied in a simple way, as illustrated in Fig. 21.3 Article bundles, which are held together in carton packages, are considered one type of autonomous objects, and customer orders are considered a second type. The article bundles are allocated to customer orders. Selection of an order is based on a ranking of all orders using adapted dispatch rules from production control, like shortest servable delivery time, or simple customer importance. Additionally, the difference between order and delivery quantities is minimised over the various product variants. This allocation influences subsequent decisions during the article distribution
Fig. 21.3 Procedure of package to order allocation
21
Limitations of Autonomous Control in Practical Applications
297
process, like e.g. selection of the next target nodes for transports of garments within the network, or storage at the factory [25]. These autonomous objects are geographically distributed between the factory and the distribution centres. For this reason, the different nodes of the transport network serve as intermediary broker elements. In these nodes information about the articles and orders is collected and disseminated. In this way, the allocation can be done for any package within the supply network. To implement the method, a hardware and software demonstrator has been designed and integrated. The demonstrator allows reading article related transponder data and allocating orders to articles. It covers the use cases ordering, packing, dispatch, warehouse entry and storage. The tool offers interfaces for order entry, RFID based identification of single garments and packages, allocation of packages to orders, and creation of packing lists. Instances of the application can be installed in several network nodes. Next to the main components, the tool includes several interfaces to RFID-interrogators infrastructure. For application at the case study, combined use of transponders on item (article) and unit (package) levels has been proposed and tested. This allows individual handling of the garment pieces, when needed. During the finishing process, the ready-made garments are equipped with easily removable transponder labels. For each piece, its type and variant data, including article number, size and colour, and an additional individual serial number referring to the unique piece is stored into the transponder label. As packages are equipped in the same way with transponder labels, they can be identified with their included articles as well. The identity of the packages can be allocated to the orders in the order database. During container stuffing, the packages are identified and counted using either mobile or gated RFID-interrogator. A packing list of all the packages and articles in the container can be automatically created and sent to the target hub. The packing list can also contain information for customs service procedures. During warehouse entry, an RFID-interrogator reads the information which is stored on the RFID label belonging to each package. The actual stock level data of the hub is updated. If an order has been modified or cancelled during transport, the packages destination will be rearranged to other customers, as described. A similar order can be brought forward for instance. If the order is still valid, no rearrangement is necessary. RFID-interrogators at the warehouse read the information of every leaving package while it is put into a container or trailer.
21.3 Discussion of Limitations The transfer of new methods into real world scenarios in practical applications is an important aspect of scientific work. The theoretically developed concepts have to be transferred to investigate the applicability and effect in real scenarios. In this context, different kinds of limiting factors lead to limitations, which lead to
298
B. Scholz-Reiter et al.
various limitations. Limiting factors reconsidered in this contribution fall into four categories: Methodical factors include applicability and performance of autonomous control
methods in industrial processes. Technical factors deal with imperfections of existing technical means and instru-
ments to implement these methods. Organizational factors, from the sociological perspective, address the effects of
autonomous control in relation to the human being. Economic factors address the balance between economic costs and benefits. This
is crucial for any enterprise thinking about adoption of autonomous control in business processes. Experiences from implementation of autonomous control show that these are the main factors, which influence the process of implementation and the success of a method in practical application. Furthermore, the factors are related to, or dependent on, each other. As Fig. 21.4 illustrates, theory on the way into practice faces different cause-effects between those factors using the implementation process. The interactions caused by effects of the factors determine the limitations of autonomous control. The limitations resulting from this interaction will be discussed in the following sections. Cause effect relationships, like the influence of technology availability on feasibility of method implementation, are finally discussed in the conclusion of this article.
Fig. 21.4 Cause and effect relationships between different limiting factors
21
Limitations of Autonomous Control in Practical Applications
299
21.3.1 Methodical Limitations Autonomous control methods include methods for modelling process flows, methods for modelling material flows, and strategies for taking decisions of autonomous objects. Process modelling methods are part of the ALEM framework [24], while material flow methods include differential equation systems, systems dynamics and event discrete simulation [20]. The strategies for autonomous controlled decisionmaking are explained in the following paragraph. Decisions, made by the logistic objects, have to be based on appropriate decisionmaking methods. Such methods generally include optimisation methods, heuristics and priority rules as well as artificial intelligence methods [30]. In particular, for routing of autonomous logistic objects in production and transport networks different methods for decision rendering have been developed. These include bio analogous methods, like e.g. pheromone based methods, rational strategies, which are either rule based, or based on estimations of future system states, like e.g. queue length estimator [20, 21]. To these, adaptations of routing mechanisms from data communication networks may be added [22]. Methodical limitations of autonomous control are here understood as those factors limiting applicability and effectiveness or performance of autonomous control. Under consideration are those factors, which stem from the nature and characteristics of the methods themselves, and not on technical, organizational or economic factors, that limit the applicability of otherwise appropriate methods. This means, that even when perfect technological implementation and human adaptation to requirements of autonomous control are assumed and economic factors are disregarded, these limitations will still be present. Methodical properties and characteristics have been already studied in some detail during the development and analysis of autonomous control methods. For instance, several of these methods, like pheromone based and QLE methods, have been compared to each other, or to conventional control methods, by simulation studies [21]. From these comparisons, limitations can be derived. These studies however have always been theoretical. To provide an additional, practical perspective, this discussion is restricted to actual experiences with methods in the mentioned projects. In both projects, decisions are mainly based on implicit knowledge and personal communication, which cannot be expressed by a general heuristic. This decision making process is difficult to be mapped into explicit rules for decision making by autonomous logistic objects. Storage management in vehicle logistics is based on daily, personal communication. The allocation will be made depending on the process requirements. Relevant processes are unloading of ships, technical treatment and disposition of vehicles. The developed decentralized decision-making process for vehicle movement at an automobile terminal refers only to the vehicles and the storage areas with the explained factors. Implicit knowledge over subsequent, probable steps of the vehicles or special characteristics of the storage areas is not considered in the theoretical method.
300
B. Scholz-Reiter et al.
In general terms, examination of disposition processes in the apparel scenario has shown that stock level and order disposition is frequently based on implicit knowledge or experiences of the disposers, often there are no explicit rules documented. Frequently there are extensive processes of personal communication between disposers of apparel suppliers and retailers for coordination (e.g. reallocation and redistribution of articles). These are based on personal trust and are frequently modified. For these reasons, it is difficult to map them into decision making rules for autonomous logistic objects. For the apparel logistics project, dispatch rules have been adapted from the field of production logistics in order to guide the decisions taken by autonomous article bundles and customer orders. The shortest servable delivery time rule mimics the slack time dispatch rule in production control [32]. Application of this rule generally tends to decrease total tardiness for customer orders. If customer based order priority is applied, garments are allocated with priority to those orders of those customers, who are of highest importance for the supplier. Customers can be ranked in importance or priority according to current or historical overall order volumes, or related criteria. Application of this rule generally decreases tardiness for large customers’ orders. However, it may increase tardiness for less important customers’ orders. Accordingly, this rule should be mainly applied in those cases, when the shortest servable delivery time rule is indifferent to priorities of two different orders [25]. These experiences confirm that one very important limiting factor in introducing autonomous control is the frequent modification of decision rules, for the logistics systems in many companies. In particular, this applies to logistic operations, which exceed company boundaries and involve business partners and customers. In these inter-company operations, conflicting goals of partners often are remedied by personal negotiation rather than application of explicit decision rules. The results of these negotiations are difficult to transform into decision rules for autonomous logistic objects.
21.3.2 Technical Limitations For the implementation of an autonomous control method, some technical equipment is necessary. Logistic objects need the ability to interact with other logistic objects within the system and to execute decisions on their own. Therefore, sensor systems as well as other ICT are of particular importance for autonomous control. Autonomous control mainly needs equipment: To get information of the actual situation and conditions, like sensor techniques
like RFID, Barcode, GPS, proximity sensor. To take decisions depending on the information; PC, micro-controllers. To communicate information on status and decisions, WLAN, GPRS, Bluetooth.
The technical limitations mainly depend on the specifications of the technical equipment. They depend on the range, operation time, precision, and computing power of
21
Limitations of Autonomous Control in Practical Applications
301
the used hardware. Based on the application of autonomous control in the scenario, different technologies lead to limitations regarding their implementation. For the apparel logistics scenario, technical requirements are less demanding than for vehicle logistics. Whereas in the vehicle logistics scenario, many different technologies for identification, localization and wireless data exchange are combined, in the apparel logistics scenario technical implementation is mostly restricted to coupling an RFID system with a software system. The main technical aspects are as follows [2]: The mandatory operation time of a computing system mainly depends on the shift
time of the work process. In vehicle logistics, we need an accumulator capacity for eight hours operation time, which is the normal shift time at the automobile terminal. Concerning this, effective power management with standby function and a customized wake up function is essential. The limiting factors regarding operation time are the size and the weight of the accumulator. In apparel logistics, transponders are merely passive, while RFID interrogators are mainly stationary, so that their energy supply is not an issue. If mobile interrogators are used, for instance for article and package identification during packing or picking, they should have a capacity for a normal shift time of eight hours as well. However, this is not a limiting factor for standard commercial products which are used in the scenario. The computing power has to provide short response times of the system. Furthermore, the computing power and the electric power are interdependent. That means for high computing power energy demand is also relatively high. This can be a limiting factor for systems where size and weight of the system are important. Normally, a bigger system can provide the necessary computing power. Barcode and RFID are common technologies for identification of logistic objects. The limitations of barcode systems often depend on external circumstances. Readability of the barcodes is complicated by raindrops of the surface or by the bleaching of the sun. Scholz-Reiter et al. (2008) define five steps of RFID technology and capability enhancement for moving RFID technology from simple identification of parts to autonomous cooperating logistic processes [26]. Intermediary steps are storage of dynamic data, decentralized data processing, communication, and finally intelligent, integrated, information based material handling. Depending on these steps, the limiting factors of RFID technology are varying. The read range and the response time of an RFID system are critical factors for the use of this technology. This depends on the one hand on the RFID system, which can work on different bands of frequency and with difference performance. On the other hand it depends on the selected transponders. A critical factor in vehicle logistics is to read a passive UHF transponder, which is placed on the outside of a vehicle’s window. In apparel logistics, cheap transponders are needed, which can be applied easily and temporarily to apparel articles, and which can easily be removed before sales to the end customer. As the exact allocation of articles to packages, and of packages to containers, has to be established, secure bulk reading of these transponders is mandatory.
302
B. Scholz-Reiter et al.
Due to the wide geographic distribution between production and distribution, standardisation between RFID systems in Asia and Europe is critical to guarantee interoperability. Positioning of logistic objects can be made by different techniques, like GPS, WLAN or RFID [26]. Depending on the requirements, rapid positioning and accuracy of the system is needed. In several cases, this is a limiting factor for a technique. For vehicle logistics, GPS was used. Due to the accuracy limits of GPS, a unique storage place allocation is not feasible. For exact allocation of vehicles to storage places, maximum variance of at most one meter is mandatory. Depending on external circumstances like the shield effect of buildings and weather conditions, GPS provides a level of accuracy between 1 and 40 m. In apparel logistics, positioning of objects is not an issue now. The application of mobile systems requires communication with other mobile systems as well as with the logistic backend system. Techniques for mobile communication are Universal Mobile Telecommunications System (UMTS) or Wireless Local Area Network (WLAN). Regarding the communication via UMTS the economical limitations are a relevant aspect concerning the fees for using the technology. A critical aspect for the use of the WLAN technology is the mandatory use of a hardware infrastructure which is affected by economic factors as well. Some technical limitations depend on economic limitations in case of necessary investments in infrastructure for the best technique or running charges. These aspects will be discussed in the chapter on economic limitations. Actually, technologies required to implement autonomous control are already available. Existing technical limitations are mainly due to individual performance, reliability and usability issues, or the non-availability of required technologies within economically reasonable cost boundaries.
21.3.3 Organizational Limitations In order to investigate limitations arising from the implementation of autonomous control in organisations, organisational culture has to be taken into account. Successful integration of new technology in an organisational context is strongly determined by organisational culture. Organisational culture guides the behaviour of employees in an indirect, unconscious and informal way, which affects the implementation of autonomous control methods. One example of, unwanted, unconscious behaviour is a refusal of passing on correct information or information, which is important for a successful implementation, to other organisation members. The reasons for this kind of behaviour can be analysed using an appropriate concept, which facilitates an in-depth study by focussing on the three levels of organisational culture – artefacts, values and basic assumptions. Therefore, the relevance to investigate organisational culture is its influence on implementation of autonomous control methods. Within the frame of changing structures, it is relevant to examine
21
Limitations of Autonomous Control in Practical Applications
303
organisational culture focussing on its ability to integrate and to adapt to new systems [33]. Some “cultures might be described as flexible or loosely integrated, which makes it easier to introduce particular changes without disturbing the whole way of life, others are so tightly integrated that change can threaten the whole” [11]. Therefore, the analysis of organisational culture is essential for the understanding of new technologies’ effects on organisations [19]. To define organisational culture, we used the definition of Edgar Schein: Organisational culture is a pattern of shared and unconscious basic assumptions, opinions, thoughts and feelings of a specific group [19]. These patterns conduct the group members’ perceptions and their interpretation of every problem they have to face. Moreover, it regulates the mental and emotional processing and the evaluation of problems. These evolutionary learned patterns enable to cope with problems of external adaptation, which implies the group’s survival, growth and adaptation to its constantly changing surrounding. Coping with problems of internal integration implies the ability to develop a system of relationships. These strategies are of proven value, obligatory for all members and will be transferred to new members. They allow dealing with requirements of their daily work so that it is possible to adapt to external challenges [19]. Besides the definition, the concept of organisational culture is chosen as well from Schein (1995). It is incorporated in the interpretative paradigm of organisational theory and is considered as an integrative concept. From the interpretative perspective, organisational culture is regarded as a system of meanings [27] and thus it is primarily close to social psychology. This concept allows analysing and understanding cultural-based problems regarding the implementation of autonomous control of two real scenarios and its available data of experience. The data, taken from these two scenarios, are based on preliminary experiences with impaired relevance for generalisation. It is not appropriate to generalize this data describing a specific organisational culture by an interpretative analysis. Therefore, the extracted know-how from the data of these two scenarios consists merely of hypotheses, which may however be helpful when considered for the implementation of autonomous control in general. This approach merely serves to raise awareness to potential upcoming difficulties. Some advices are given in Chap. 4. From the perspective that culture is a system of meanings, Schreyögg (2004) and Zielowski (2006) assert that distinctive patterns of perception and interpretation strongly develop and influence the behaviour and the attitudes of the group members and the group’s operational system [27, 33]. Therefore, by analysing the reasons for limitations using Schein’s concept (1995), we try to examine how people perceive, categorise, believe and value an upcoming structural change caused by autonomous control methods. Schein (1995) maintains that all these mentioned cognitive aspects are based on and conducted by a system of basic assumptions, which leads to stability of and meaning in organizational culture [19]. Schein (1990) asserts that “once one understands some of these assumptions, it becomes much easier to decipher the meanings implicit in the various behavioural and artifactual phenomena one observes” [17].
304
B. Scholz-Reiter et al.
The mentioned system of values and assumptions can be meaningfully applied towards technology. Use of specific and proven technologies can build up stability and enable the group to distinguish itself from others and form its identity [18]. From this perspective, initial refusing behaviour of employees is considered a predictable expression towards technological changes, in order to prevent an irritation of basic assumptions and values associated with well-established technologies. Existing values and assumptions have to be redefined, if new technologies replace well-established and proven technologies [18]. Therefore, refusing behaviour against implementation is a psychological reaction against the irritation of the groups’ self-perception and identity. Indeed, this is not the only reason for resistance against changes such as the implementation of autonomous control. By reason of data privacy, it is not allowed to present the interpretation of data of the two mentioned scenarios in detail. Thus, the following section merely touches the surface possible of reasons for employees refusing behaviour regarding one of the two mentioned real scenarios. In the scenario of vehicle logistics, the organisation has experienced a process of change, which provoked insecurity and sceptic attitude, before the implementation of autonomous control had started. Ambivalent attitudes towards the implementation may be caused by a merger of different divisions, due to a strategic restructuring of the organisation before the implementation. The affected divisions prevent further intervention in the structure via implementation of autonomous control, because their several group identities and self-perceptions are already menaced by the merger. Moreover, this structural change is accompanied by rationalisation and wage reduction, which leads to a feeling embossed by insecurity regarding the continuity of the group and every individual. According to Schein (1995), groups impede further structural change and fight for their survival through refusing behaviour [19]. Regarding this, disguising information needed by researchers to adapt autonomous control methods is a comprehensible reaction to preserve their group’s continuity. Moreover, if the implementation of autonomous control is interpreted as a rationalisation, which is more or less a possible interpretation caused by logical deduction from the method’s function, the employees of these divisions will feel menaced by this intervention. It reduces the need of the employees’ competence in specific areas of responsibility. Furthermore, the economic crisis of the year 2009 influenced this cultural phenomenon of refusing behaviour as well. The members of an organisation may consider the economic crisis as a factor of insecurity, which intensifies collective feelings of mistrust and anxiety. It also increases the likelihood of rationalisation, which leads to turmoil in organisations, as well as an intervention in the organisational culture by the implementation of autonomous control methods. In the apparel logistics scenario a structural change such as a transition of leadership occurred during the implementation. Such a change leads to scepticism towards implementation of autonomous control methods as well. Thus, concerning an imminent or on-going transition, it is expedient to check out the culture’s ability to face implementation and to examine the possibilities to adapt to and to integrate autonomous control methods and its appropriate technologies.
21
Limitations of Autonomous Control in Practical Applications
305
21.3.4 Economic Limitations Economic problems address the balance between costs and benefits, which is crucial for any enterprise thinking about adopting autonomous control in its business processes. As mentioned above, economic limitations depend on the already discussed methodical, technical and organizational limitations. That implies that monetary efforts, for example the adaptation of a method to a scenario, or development and acquisition of new technology and equipment, can prevent limitations for those other factors. Otherwise, less investigation creates limitations in some factors. A technical limitation in the field of vehicle logistics is the precision of localization which can be achieved. A precision of one meter is only feasible with a certain infrastructure. The procurement and implementation of this technology is a crucial economic factor. A project with the aim of implementing autonomous control has wide range effects. Usually, the implementation of autonomous control is not possible to be confined to only a part of a process. Hence, it is a complicated and extensive process involving uncertainties and inaccurate cost planning. Furthermore, autonomous control has not been implemented in business processes yet. Thus, there are no previous experiences available, which actual projects can build on. The technical data from other ICT implementation projects serve as basis for estimating hardware costs in this case. However, no data is available for estimating the influence caused by the change of the control method. No previous experience exists so far to serve as reference data on how to estimate costs of implementing autonomous control, in particular less tangible costs, such as those for small error corrections during initial process implementation. Therefore, the uncertainty of the cost data is relatively high. One important factor for economic limitations is the length of the pay-off period. For IT-projects, like the implementation of autonomous control, a pay-off time of two years has to be realized. A crucial economic factor for vehicle logistics is the number of vehicles that are identified by a smart label from the factory. The explained process of autonomous controlled decision making at an automobile terminal can only be economically rewarding, if nearly all vehicles are identified by a smart label by the factory. In apparel logistics, the number of objects, which have to be manipulated, is large, whereas the value of one object is limited. This means that only very cheap technologies can be applied economically to the individual logistic object. For an apparel article, only a cheap passive transponder of a few cents worth can be applied economically. This causes a very serious constraint on the feasible degree of autonomous control, as information processes, and therefore decision-making, has to be implemented completely separated from those objects in a separate software system. In summary it can be said, that economic limitations are the most important decision-making basis for any implementation project. The balance between the investigation in the different factors, like technique or the organisational structure, is highly relevant for the successful implementation of autonomous control.
306
B. Scholz-Reiter et al.
Economic investments can support various sectors and time horizons in the implementation process of autonomous control. Various sectors means that a high volume of investigation in the applied technology can provide the best technical equipment, but if there is no investigation in the organisation, like workshops to explain the technology to employees and the work councils, the possibility of rejection is relatively high. Investigations have different time horizons. Mostly, short-term investigations react on a crucial problem long-term investigations often prevent crucial problems.
21.3.5 Limitations in the Scope of the Practical Scenarios In the prior sections four types limiting factors have been explained, and some examples for the scenarios have been given. The following table gives a structured summary of the specific limiting factors experienced in the scenarios (Table 21.1). This summary is not a complete list of all limiting factors experienced in both described scenarios, but it can provide a quick overview of the most important factors.
21.4 Conclusion A purpose of this article was to report on experiences gained during two projects on limitations of autonomous control. Thus, a dedicated practical perspective was adopted. The available data basis of the mentioned projects is not comprehensive enough to make conclusions which are universally valid. Nevertheless, the outlined experiences should be helpful to realise additional projects, which transfer autonomous control into industrial practice. The results show that it is necessary to address limiting factors in an organized manner during such projects. The relationships between the mentioned aspects concerning the implementation of autonomous control have to be understood in their complex interaction to find appropriate strategies to prevent or reduce them. Limiting factors can be related as positive or negative cause and effect relationships. Positive cause and effect relationships can support the implementation of autonomous control. Negative cause and effect relationships between limiting factors should be avoided or prevented as far as possible. Some examples of limiting factors and advices to face them can be provided in the following paragraphs. For the relationships between methodical, technological and organisational factors some examples are provided. Economic factors, as the basis for the final decision in industrial investment, were discussed in addition to each of the other relationships. The developed autonomous control methods must be adapted to the environmental circumstances to ensure a successful implementation into a real world scenario. Furthermore, the available technological infrastructure must be suited to the
21
Limitations of Autonomous Control in Practical Applications
307
Table 21.1 Summary of limitations in the practical scenarios Limiting factor Vehicle logistics Apparel logistics Methodical The expression of implicit Decisions in storage manage- Stock level and order disposition is often based on experience of knowledge in general ment depend on daily perdisposers or implicit knowledge; heuristics sonal communication, which decision rules are frequently can be only insufficiently modified due to individual cusmapped in general heuristics tomer preferences Coordination between suppliers and retailers depends on personal communication based on personal trust Technical The technical requirements Operation time of 8 h Secure bulk reading of transponbased on the requirements ders of the scenario Robustness and shock resis- Standardisation of RFID systems tance against dropping down in different regions (Asia and Europe) RFID-Reader response time < two seconds Positioning accuracy up to four meters Organisational Scepticism towards implemen- Scepticism towards implementation of autonomous control tation of autonomous control caused by caused by Merger of different functions Scepticism towards autonomous control during transition of leadership Employee’s fear to lose criti- Unclear economic prospects durcal qualifications ing economic downturn Unclear economic prospects during economic downturn Economic Balance between cost and Cost estimation for imple- Cost estimation for implemenbenefits mentation of autonomous tation of autonomous control is control is difficult difficult Every vehicle must be Large number of objects, limited equipped by a smart label to value of each single object be identified Cheap, easily applicable transponders on all items and packages
requirements resulting from the method. Otherwise, the non-available technology prevents the success of a method. Technical limitations emerge from economic factors. Investment in technological equipment can benefit the success of the technical implementation. In most cases, less investigation into a technology generates problems with the performance of the technology. Additionally, investment in increased
308
B. Scholz-Reiter et al.
usability of a technology prevents effects of limiting factors. The economical benefits of a method’s application are examined by simulation studies in fundamental research. In these studies development of logistic key factors like lead times, set up time or workload is investigated. The results of these studies allow estimating the economic feasibility of the methods. Based on the economic feasibility, decision makers in the industry can trigger implementation projects. The performance of the methods influences the limiting economic factors immediately. Autonomous methods may require new equipment, for instance, for equipping each logistic object with its own sensors or information processing capacity. The required technology can be entirely lacking, be principally available but too deficient in capability, or simply too expensive. The maturity of technology should be appropriate to the organizational culture’s stage of development, which also includes the structure of an organization. An organization, which is experiencing structural changes, should be spared from ill-conceived technology for implementation during this period, because it would threaten the organizational culture in addition to its previous change. This dual burden could impede a successful implementation due to organizational culture’s resistance against further interventions in its culture. This would lead to a capital investment for a new technology without early possibility for a successful implementation. Apart from matured technology, the implementation of autonomous control methods requires the introduction of the new technical equipment. The executive employees have to be able to handle the equipment. Therefore, the executive employees have to be trained with the new equipment in time. Moreover, they have to comprehend the imperative of this new technology for the organisation’s success on the market. This comprehension is necessary to connect a deeper sense to the use of the new technology and to associate a certain meaning with its function in an organizational context. However, considering the disposition of meaning and sense might be important for new technologies’ integration into organisations’ self-perception. According to Markus and Robey (1988), autonomous control methods’ relationship to organisations could also be seen from a perspective, in which “the uses and consequences of IT emerge unpredictably from complex social interactions” [34]. Thus, some of the meanings, which are associated with autonomous control methods, are subjective as well as socially constructed. Regarding a successful implementation, it is important to investigate the subjective meaning of autonomous control in specific organisational context. The reason is that different specific organisational cultures percept, interpret and evaluate autonomous control methods in cultural-specific ways. Thus, any specific organisational culture should be considered or investigated in advance to be prepared for a successful implementation. Moreover, to reduce upcoming cultural resistance, an implementation should be slowly introduced and established by leadership involving its manager and executive employees. This process may be supported by a specialized consultancy, which may be engaged to check out the ability of organizational culture for integration and adaptation of autonomous control in advance. Furthermore, it may examine an appropriate strategy for implementation including the basic assumptions of organisational culture. This approach is the basis for preparing a successful strategy for implementation considering possible cultural reactions. Without
21
Limitations of Autonomous Control in Practical Applications
309
an investigation and preparation of organizational culture in advance, cultural reaction could extend implementation in duration and costs. According to Schein (1990), a change of members’ behaviour towards implementation of new technologies does not create high costs for the organisation unless consultancy attempts to change organizational culture [17]. McCarty (1989) assented that it may be more appropriate to use the creative potential of the pre-existing culture, rather than attempting to change culture [11]. The attempt of changing a culture will lead to high costs, if it extends into a long term project depending on organisational cultures complexity and history. Economic limitations strongly influence the implementation of new technologies. For instance, if a company does not invest in the preparation of the organisational culture in advance, for example by checking out the applicability of new technology in this organizational culture’s context, the organisational culture will strongly impede this project. Organisational change should be able to deal with organisational resistance, otherwise new anxieties and aggression emerges from the experienced situation [11]. Regarding this aspect, there is a need to prepare organizational culture for implementation to prevent strong anxieties in advance. The explained cause and effect relationships are only a selection of the entire range of relationships, but they can give a first insight into the problem of different, each other relating aspects caused by limitations of autonomous control.
Online document COMMISSION OF THE EUROPEAN COMMUNITIES (2009) Internet of Things - An action plan for Europe. Europa. http://ec.europa.eu/information_ society/policy/rfid/documents/commiot2009.pdf. Accessed 26 August 2009.
References 1. Biermann M (1994) Paradigmatische positionen in der organisationsforschung und ihre bedeutung für eine organisationskulturforschung. Dissertation. Repro Center, Ibbenbühren 2. Böse F, Windt K (2007) Autonomously controlled storage allocation. In: Hülsmann M, Windt K (eds) Understanding autonomous cooperation and control in logistics – the impact on management, information and communication and material flow. Springer, Berlin, pp 351–363 3. Böse F, Windt K (2007) Business process modeling of autonomously controlled production systems. In: Hülsmann M, Windt K (eds) Understanding autonomous cooperation and control in logistics – the impact on management, information and communication and material flow. Springer, Berlin, pp 73–83 4. Böse F, Windt K (2007) Catalogue of criteria for autonomous control in logistics. In: Hülsmann M, Windt K (eds) Understanding autonomous cooperation and control in logistics – the impact on management, information and communication and material flow. Springer, Berlin, pp 57–72 5. Bruckner A, Müller S (2003) Supply chain management in der bekleidungsindustrie. Forschungsstelle der bekleidungsindustrie, Köln 6. Burrell G, Morgan G (1979) Social paradigms and organizational analysis-elements of the sociology of corporate life. Heinemann Educational Books, London
310
B. Scholz-Reiter et al.
7. Bussmann S, Jennings N, Wooldridge M (2004) Multiagent systems for manufacturing control: a design methodology. Springer, Berlin 8. Dyer WG (1986) Culture change in family firms: anticipating and managing business and family transitions. Jossey-Bass, San Francisco 9. Hülsmann M, Windt K (eds) (2008) Understanding autonomous cooperation and control in logistics. Springer, Berlin 10. Kieser A (ed) (1999) Organisationstheorien. Kohlhammer Verlag, Stuttgart/Berlin, Köln 11. McCarthy E (1989) The dynamics of culture, organisational culture and change. AI & Soc 12(3):155–184 12. Orlowski WJ, Robey D (1991) Information technology and the structuring of organisations. Inf Syst Res 2(2):143–169 13. Overmeyer L, Nyhuis P, Höhn R, Fischer A (2006) Controlling in der intralogistik mit hilfe von pre processing labels. In: Pfohl H, Wimmer T (eds) Steuerung von logistiksystemen – auf dem weg zur selbststeuerung. Dt. Verkehrs-Verlag, Hamburg, pp 205–215 14. Pfohl HC (2000) Logistiksysteme – betriebswirtschaftliche Grundlagen. Springer, Berlin 15. Rügge I, Ruthenbeck C, Scholz-Reiter B (2009) Changes of HCI methods towards the development process of wearable computing. In: Kurosu M (ed) Human centered design. Springer, pp 302–311 16. Sampler JL (1996) Exploring the relationship between information technology and organisational structure. In: Earl MJ (ed) Information management: the organisational dimension. Oxford University Press, Oxford 17. Schein EH (1990) Organizational culture. Am Psychol 45:109–119 18. Schein EH (1991) Organisationskultur: Ein neues unternehmerisches Konzept. In: Dülfer E (ed) Organisationskultur: phänomen, philosophie, technologie. Poeschel, Stuttgart, pp 23–37 19. Schein EH (1978) Career dynamics. Matching individual and organizational needs. Reading, MA 20. Scholz-Reiter B, Böse F, Jagalski T, Windt K (2007) Selbststeuerung in der betrieblichen Praxis – Ein framework zur auswahl der passenden selbststeuerungsstrategie. Indus Manag 23(3):7–10 21. Scholz-Reiter B, Görges M, Philipp T (2009) Autonomously controlled production systems – influence of autonomous control level on logistic performance. CIRP Ann Manuf Technol 58(1):395–398 22. Scholz-Reiter B, Rekersbrink H, Freitag M (2006) Kooperierende routingprotokolle zur selbststeuerung von transportprozessen. Ind Manag 22(3):7–10. GITO-Verlag, Berlin 23. Scholz-Reiter B, Rekersbrink H, Görges M (2010) Dynamic flexible flow shop problems – scheduling heuristics vs. autonomous control. CIRP Ann Manuf Technol 59(1):465–468 24. Scholz-Reiter B, Sowade S, Hildebrandt T, Rippel D (2010) Modeling of orders in autonomously controlled logistic systems. Prod Eng Res Dev 4(4):319–325 25. Scholz-Reiter B, Teucke M, Schweizer A, Özsahin M-E (2010) Flexible product allocation in distribution processes in an apparel supply chain. In: Proceedings of the 2010 international conference on logistics and maritime systems (LOGMS 2010), Busan, 15–17 Sep 2010, 604– 613 (on CD) 26. Scholz-Reiter B, Uckelmann D, Gorldt C, Hinrichs U, Tervo JT (2008) Moving from RFID to autonomous cooperating logistic processes. In: Miles SB, Sarma SE, Williams JR (eds) RFID technology and applications. Cambridge University Press, Cambridge, pp 183–197 27. Schreyögg G (2004) Unternehmenskultur zwischen stabilität und wandel. In: Manske F, Lange H (eds) Kultur im Veränderungsprozess: Kultur als analytische Kategorie in der arbeits- und organisations-, der Innovations- und der umweltforschung. Nomos Verl.-Ges., Baden Baden 28. Ten Hompel M, Trautmann A (2006) Modellierung prozessadaptiver agenten zur steuerung autonomer lagerfahrzeuge. In: Pfohl H-C, Wimmer T (eds) Steuerung von logistiksystemen – auf dem weg zur selbststeuerung, 40–54. Deutscher Verkehrs-Verlag, Hamburg 29. Vallejo MC (2008) Is the culture of family firms really different? a value-based model for its survival through generations. J Bus Ethics 81:261–279
21
Limitations of Autonomous Control in Practical Applications
311
30. Weigelt M (1994) Dezentrale produktionssteuerung mit agenten-systemen. Entwicklung neuer Verfahren und vergleich mit zentraler lenkung. Deutscher Universitäts-Verlag (Gabler Edition Wissenschaft), Wiesbaden 31. Windt K, Philipp T, Böse F (2008) Complexity cube for the characterization of complex production systems. Int J Comput Integrated Manuf 21(2):195–200 32. Kistner K-P, Steven M (2001) Produktionsplanung. Physica-Verl, Heidelberg 33. Zielowski C (2006) Managementkonzepte aus Sicht der organisationskultur: Auswahl, Ausgestaltung und Einführung. Dt. Univ.-Verl. Wiesbaden 34. Markus ML, Robey D (1988) Information technology and organizational change: causal structure in theory and research. Manag Sci 34(5):583–598
Chapter 22
Autonomous Control in Production Planning and Control: How to Integrate Autonomous Control into Existing Production Planning and Control Structures Marius Veigt, Farideh Ganji, Ernesto Morales Kluge, and Bernd Scholz-Reiter
22.1 Introduction The changing conditions of present markets lead to an increasing complexity and dynamic of production logistics. Conventional production planning and control structures and methods which are characterized by central planning and control processes, cannot meet these new requirements. This is due to the structures and methods of central planning which do not allow fast and flexible adaption to changing environmental influences [1]. The new challenge for production planning and control (PPC) systems is the reaction to changing environmental influences in real time [2]. An opportunity to meet these requirements is the introduction of autonomous control. In this approach autonomous decision functions are shifted to single objects. Due to this splitting of the decision function the possibility arises to handle the mentioned complexity and to react immediately to changing environmental influences. Existing structures for production planning and control are organized centrally, however, whereas the structure of autonomous control is decentralized. This contribution outlines structural changes arising from the integration of autonomous control in production planning and control. One of the main questions is, on which layer of the PPC-process autonomous control can be introduced. Another question is, how autonomous control can interact with existing PPC-systems and where there are obstacles and limitations for its implementation. This contribution is structured as follows: Chap. 2 gives a summary of how production planning and control is structured nowadays and of what kind of information systems are used for PPC. Chapter 3 describes technologies enabling autonomous control. In Chap. 4 an exemplary solution of an autonomously controlled factory is presented which was developed in the Collaborative Research Center 637. Finally
M. Veigt (B), F. Ganji, E.M. Kluge, and B. Scholz-Reiter Department of Planning and Control of Production Systems, BIBA, University of Bremen, Bremen, Germany e-mail:
[email protected] M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6_22,
313
314
M. Veigt et al.
Chap. 5 summarizes the potentials as well as the limits of an implementation of autonomous control in PPC.
22.2 State of the Art of PPC Structures This section describes how production planning and control is organized nowadays. Therefore the automation pyramid, shown in Fig. 22.1, will be used. This model divides the planning and control structure of an enterprise into six layers. In practice there are no clear dividing lines between the layers but this model provides the opportunity to differentiate challenges and limits for production planning and control on clearly divided layers. The top layer is determined by the strategic planning function of an enterprise. There is no direct intervention in production planning and control processes from this layer [2]. Consequently, corporate management is not the focal point of this contribution. On the manufacturing management layer customer orders and procurement are managed. Furthermore, on this layer the production planning starts by creating a rough production plan and providing this information to the lower layer [2]. On layer 3 (production management) the rough production plan is specified. This means, the scheduling of orders (e.g. for a day, shift or hour) and the disposition of workers and machines are worked out. According to the schedule, the orders are released for production processing. Moreover, finished orders are reported to layer 4. On layer 2 information, e.g. how to handle a part, must be processed promptly. Here, information is distributed to workers and machines to complete the production process. The information which is provided for machines must be forwarded to layer 1 to control the machines. On this layer information is restructured and transmitted to actors on layer 0. The actors affect physical processes, and sensors monitor the processes and send a feedback to layer 1.
Fig. 22.1 Automation pyramid for PPC
22
Autonomous Control in Production Planning and Control
315
The layer structure in Fig. 22.1 is theoretical. In practice some layers may not exist or there are further intermediate layers. Furthermore, there is no clear dividing line between the layers [2].
22.2.1 Information Systems for Production Planning and Control From layer 3 to layer 5, Enterprise-Resource-Planning-Systems (ERP-Systems) have been established. In some cases the ERP-System is complemented with an Advanced-Planning and Scheduling-System (APS-System). ERP-Systems manage a multitude of enterprise information in a central database and provide several administrative and economic functions. At this point, production planning starts by creating a rough production plan. Manufacturing-Execution-Systems (MES) are used on layer 2 and 3. An essential function of MES is production planning and control. On layer 3 MES generate a detailed production plan as well as controlling and monitoring the production. On layer 2 several MES control a production process. The structures of MES are historically grown, whereby the MES are often differentiated by functional aspects, e.g. warehouse management system and material flow system. MES provide information for controlling the machines. On the automation layer this information is restructured to be applicable for controlling the machines by a programmable logic controller (PLC) etc. The controller uses actors (A) to initiate physical processes and collects data from sensors (S). Figure 22.2 outlines this information system structure, following [1, 2].
Fig. 22.2 Information and control systems for PPC
316
M. Veigt et al.
22.2.2 Functions and Limitations of the Information Systems This section will detail the functionalities and limitations of information systems of production planning and control, such as ERP, MES as well as control systems on the automation layer.
22.2.2.1 ERP and APS Enterprise-Resource-Planning-Systems have many uses for managing and administrate the enterprise’s resources as well as for optimizing business processes. Hereby ERP-Systems provide a range of functions for diverse departments of a company, e.g. purchasing, production, sales, quality management, accounting etc. [3]. However, the comprehensive usability of ERP-Systems is the reason why the systems are not made for special tasks. These systems are inapplicable especially for production planning and control. Points of criticism are that planning values are estimated with no reference to actual workloads [4] and the planning does not consider the capacity limit of the production [5]. Another point of criticism is that the planning takes a long time for complex production systems and due to this, short-term events cannot be considered during the planning period [1, 6]. This means that the planning is inflexible when it comes to rush orders, changes in demand or in time [1]. Advanced-Planning and Scheduling-Systems were developed to eliminate these weak points and to enable planning along the entire supply chain. APS-Systems receive data from ERP-Systems [6] and generate production plans faster than ERPSystems by using heuristics. However, even APS-Systems cannot provide a solution for a detailed production planning and control with the possibility to react to shortterm events [1, 4].
22.2.2.2 MES Manufacturing-Execution-Systems are made for production planning and control. These systems can be combined with ERP-Systems. While in ERP-Systems the economic functions are implemented, MES provide functions for production planning, controlling and monitoring [2, 7]. MES can exchange data with ERP-Systems, such as working plans. They are able to store data or customer orders. Hereby redundant data storage can occur, e.g. administration of stocks [7]. Besides, linking problems can arise between MES and ERP-System as well as between the diverse Manufacturing Execution Systems [2]. Due to this problem the idea exists to equip MES with adapters and to link these systems by using a middleware [7]. Another idea is a horizontal integration on the MES-level. This means that all MES-functions should be integrated in one Manufacturing Execution System [6]. By using MES, production can be planned and controlled in more detailed and is more flexible than by using an ERP- or APS-System. But an optimization of the production process can only be achieved by using an offline-optimization based on
22
Autonomous Control in Production Planning and Control
317
simulation models. The use of operating data to react immediately on events has not been realized yet.
22.2.2.3 Automation On the automation level, control systems must be real time capable to react within milliseconds [2]. On this level, programmable logic controller (PLC) or numerically-controlled machines (NC machines) have been established. These systems can fulfill the requirements on this level because they are real time capable, fault-tolerant and have a high technical reliability [2]. These systems receive commands from upper levels and distribute these commands to actors, such as motors or hydraulic cylinders, in order to start the demanded physical process. Furthermore, the process is monitored by sensors. Sensors can be scales, thermometers, light barriers etc. A function of sensors highly important for logistics is to identify orders and objects, e.g. parts, in the production area. Therefore different identification technologies are used in industrial environments. The barcode, like the EAN-Code, is very common and provides a low cost solution for automated identification (cf. [8]. Another identification technique is Radio Frequency Identification (RFID). The advantages of RFID in comparison to the barcode are the possibility to identify objects without line of sight and to write and change data on the memory of the RFID-transponder. Furthermore the data storage of a RFID-transponder can be much bigger than that of a barcode label [9].
22.3 Technical Approaches for Autonomous Control Today’s complexity of computer based control systems causes high cost for their development. This is the main reason for the adoption of heterarchical and decentralized manufacturing control systems [10] as well as the focus on rising flexibility in manufacturing systems for mixed-model-assembly. Furthermore, this approach affords fault-tolerant handling as well as flexibility in reacting to context changes. Hierarchical and centralized manufacturing systems, in contrast, are inflexible, i.e. they are not able to react to changing production styles and highly dynamic variations in product requirements. The reconfiguration capabilities of the manufacturing systems are limited and are highly cost intensive [11]. Particularly traditional integrated information systems such as ERP systems are complex and thus expensive to adapt [12]. In order to establish autonomous logistic systems we imply technologies that offer high flexibility for control systems. These technologies are a prerequisite for physical items and corresponding control systems. Considering distributed and modular design for autonomous control systems, it is necessary to merge the physical world with the control system. Autonomous systems focus on intelligent logistic objects that are able to control themselves in production processes. Autonomous
318
M. Veigt et al.
objects possess the ability of unique identification, of sensing their environment, of communicating with other objects and finally of making their own decisions. To manifest these abilities they depend on physical as well as computerized configuration. On the physical side there is the necessity of adding mobile data devices to the physical object. As a consequence, parts of the control system have to be distributed to the physical objects and have to be specific for each intelligent object type. Having the mentioned requirements in mind the autonomous control at the MES level can be applied independently from the higher information levels such as the production management system. In this chapter we describe technical approaches focusing on autonomous process planning.
22.3.1 Unique Identification and Localization One important feature of autonomous control is the specific identification possibility of every single object. Today, this characteristic can be achieved with the help of RFID technology. The embedding of RFID devices into physical objects as a permanent part of these objects is an answer to the long-term possibility of unique identification along the supply chain. There are first approaches to integrate RFID devices into physical objects such as casting metal parts [13]. Another capability of RFID technology is keeping processing data within the objects themselves. So it is possible to communicate the data in a new environment without connecting to a central data base system. Thus it is necessary to use the standardized electronic product code (EPC) developed by the international organization EPC global. In addition, the unique identification of objects allows their localization within the production shop. The tracing of products will be also possible by using the same RFID technology for registering freights for transport or by entering a new production or storage area. In summary the RFID technology is an important tool for the realization of autonomous logistics even if further development in this area is absolutely essential.
22.3.2 Agent Based Control Agent based control systems have already been adopted to enhance existing manufacturing information management systems. The abundant research on multi-agent technology since the 1970s has been motivated by the basic requirements of modern manufacturing systems [11]. For implementation of autonomous control, the approach of decentralized and modular decision making processes directly on the object level, also motivated the deployment of multi-agent systems for manufacturing control [14]. The question of agent behavior is the central point of implementation. Basically, every kind of software is suited for the realization of autonomous control if it implements the same behavior of agents [15]. The facilities given by standardized protocol and communication languages are beneficial
22
Autonomous Control in Production Planning and Control
319
and favor the deployment of multi-agent systems. Agents in a multi-agent environment interact in a standard way defined by FIPA [16] which is a subsection of IEEE since 2005. In particular, the format and semantics of messages (Agent Communication language, ACL) sent between agents are defined by the FIPA standards. Also the protocols for certain interaction processes based on the speech act theory [17] are standardized by FIPA [16]. These multi-agent systems are suited for Information management in manufacturing systems, but also for shop floor-level process management [18, 19]. Apparently software agents meet the requirements of modern and complex manufacturing systems, considering their characteristics such as autonomy, social ability, responsiveness, proactiveness, adaptability, mobility, veracity, and rationality [20]. The prerequisites to meet these requirements are for example the necessity for distributed organization, scalability, failure-tolerance and integration of humans in soft- and hardware control systems [11]. In manufacturing logistics agents are representatives of different logistic objects, e.g., products, assembly machines or hardware control items. According to the above mentioned requirements, there are three different architectural categories for agent based manufacturing systems: the Hierarchical approach, the Federation approach and the Autonomous Agent approach [21]. Autonomous agents have characteristics that are necessary to achieve a completely autonomous control system. They are not controlled or managed by any other agents or human beings, but communicate directly with other agents or humans. They have knowledge about their environment and have their own goals and associated set of motivations [22]. Considering these characteristics the Autonomous Agent approach seems to be the best suited approach to develop autonomous and intelligent logistic systems. The decentralized and modular design of both systems allows horizontal as well as vertical exchange of information between planning management and process management systems of enterprises.
22.3.3 Autonomous Methods and Decision Algorithms In this paper we focus on flow shop manufacturing systems that are designed for a mixed-model-assembly. The developed autonomous methods require the merging of material flow system and information flow between the intelligent objects. The approach of enabling products to act autonomously in a manufacturing system leads to the development of methods for decision making which are processed at the virtual side of a physical object [23], assuming that a logistic object consists of a physical part and an (virtual) agent. Decision-making methods have to be implemented, for example, inside the agent behaviors in the case of a multi-agent system approach. The flexibility potentials can be picked up by means of autonomous methods: firstly by dynamic changes of product variants regarding production volume and customer order and secondly by dynamic changes of production or assembly step order. To make decisions based on the mentioned factors or on unexpected faults, the products require a real-time information exchange with both, physical
320
M. Veigt et al.
structure and production planning system [24]. There is some static environment information, for example, about possible routes between manufacturing machines or the abilities of the manufacturing steps of each machine. Such information can be communicated as environment configuration once only. Dynamic changes such as new customer orders have to be communicated immediately. An unexpected fault has also to be registered at once. The decision-making process regarding product variants or the next possible manufacturing steps is basically only a part of the autonomous control of an object. After a new decision is certain, for example a product has to actuate the accordant material flow system to move itself to next destination. There are also some autonomous routing methods to find the shortest way to reach the next goal. The developed methods are based on planning and optimization heuristics that are already known from central control systems, for example, genetic algorithms, neural networks or fuzzy logic [25].
22.4 Autonomous Control and PPC This section focuses on the integration of autonomous control in existing PPCstructures. The first step is the question on which layer the autonomous control should be integrated (4.1). The second step is how this integration should be implemented (4.2).
22.4.1 Integration of Autonomous Control in Existing PPC-Structures Autonomous control is based on the distribution of centralized complex decisionmaking in existing structures. Thus even a complex system can react immediately on dynamical influences. On layer 4 and 5 the decision making is very complex (see Fig. 22.3). This is caused by several external and internal influences affecting the decision process and also the consequences which a decision on this layer has to decisions on lower layers. However, the decision making on layer 4 and 5 is strategic and long-termed and this means it is not time critical. Hence, the introduction of autonomous control on these layers seems to be only little profitable. On layers 2 and 3 the operational production planning and control is performed. On these layers the decision making is also complex. Furthermore a reaction on changing environmental influences in real time is required [2]. Nowadays a couple of control systems are used to reduce the complexity on these layers, e.g. for production planning, for warehousing, etc. By using autonomous control this complexity can be handled. As it is shown in Fig. 22.3 layer 2 and 3 can be merged into one layer. The decision-making process contains the detailed production plan, production control and warehousing as well as coordination of the material flow. Although this system is very complex, the decisions can be taken immediately by using the distributed decision making of autonomous control.
22
Autonomous Control in Production Planning and Control
321
Fig. 22.3 Integration of autonomous control in existing PPC-structures
In comparison to the upper layer, the decision making on layer 1 is simple. On this layer a control system has to react in micro- or in milliseconds. Due to the simplicity of the decision making nowadays rule-based control systems (such as PLC) are used on this layer and are very reliable on the technical side. Therefore an introduction of autonomous control on this layer does not provide added value. The following structural model shows the suggested place of autonomous control in the existing PPC-structure.
22.4.2 Autonomous Control as a Manufacturing Execution System The main benefit of using autonomous control can be achieved if the complexity of decisions and the dynamic of changing environment influences are high. As a consequence autonomous control seems to be very beneficial when it is used as a Manufacturing Execution System. To design autonomous control, we require a virtual model of the physical system. This model has to contain all objects relevant to the modeled system and has to represent the current state of them. Therefore, it is an important requirement to have a continuous and real-time communication with the real world [24]. In this manner the autonomous objects (for example finished products or parts) will be able to autonomously allocate resources, such as machines or means of transport, and to coordinate their production process themselves. With the approach of autonomous
322
M. Veigt et al.
agents within a multi-agent system it is possible to assign an agent to any involved object in the real world. The automation level contains material flow system, storage and machines in case of automated production. Today’s established control system for these elements, e.g. Programmable Logic Controller (PLC), could be even integrated in the model of autonomous control. The problem of interacting between these elements of the automation level is solved by implementing agent messaging. The expansion of autonomous behavior to the higher ERP-level depends on the integration possibilities of traditional ERP-systems in the developed virtual model of the autonomous control. In order to do this, such systems should be able to make their information available and also to process the backflow of information by means of interfaces. In this case, they could be integrated in the virtual model of the whole system, for example by being represented by one (or more) agent. Apart from centralized production systems, autonomous products or parts have the role of decision makers and arrange their own transport processes by themselves using applicable autonomous control and routing methods. For this purpose they need information about their environment. Furthermore, they have to update their knowledge about the current state of their environment permanently. This information concerns for example dynamic changes of current machine, storage and transport availability and also remaining numbers of customer orders. In summary, there is a necessity of knowledge management for autonomous agents and decision makers. Remaining on the MES-level, the available information about current changes is limited to the information about this level only. However, flexible allocation of available resources (e.g. machines) is still the main focus. For example, the products can receive messages about a machine fault and avoid choosing machines that are out of commission. Flexible change of the final goal (product variant) is indeed only based on past information relative to the beginning absolute number of customer orders.
22.5 An Autonomous Assembly Line In this chapter we focus on the first integration attempts of intelligent products in the context of autonomous logistics processes. We present a demonstrator of decentralized control in an industrial material flow system with autonomous control methods. Thereby, the decision competencies of planning and control are shifted to logistic objects by using a multi-agent system. The implemented hardware demonstrator called Factory of Autonomous Products [26]. The intention to reflect the ongoing work on developing autonomous control methods in logistic systems, specifically in manufacturing logistics, is the main motivation for this development. The following scenario shows how we investigate the feasibility of the methodological research results originating from the Collaborative Research Center (CRC 637). We illustrate an autonomous assembly system for automotive tail-lights to investigate its applicability in the domain of manufacturing logistics. In the upcoming sections we will describe the assembly scenario,
22
Autonomous Control in Production Planning and Control
323
the modeling of distributed autonomous control system based on multi-agent system JADE (Java Agent Development Toolkit), the deployed hardware, such as the material flow system and RFID technology, and finally the autonomous control and decision making methods.
22.5.1 The Scenario The scenario presents an assembly system for automotive tail-lights with six processing stations. Every station can be configured for one or more assembly steps, whereby the assembly itself is designed to be a manual task. The system is a mixed model assembly and is tailored to manufacture three type variants of the same product (automotive tail-light with colored, clear or dark diffuser). The assembly process is composed of five stages including the starting station implemented as input/output for the material flow system (Fig. 22.4). Following the idea of coupling the physical with the virtual world via agents, the metal cast parts will be the intelligent products that act as decision makers in this scenario. According to their own decisions which variant to target they are also capable to schedule the step order of their assembly process considering existing logical constraints. Close to reality, intended malfunctions or failures to the system are applied. It is also possible to take influence on the entire situation by changing the customer orders.
Fig. 22.4 Assembly scenario with five stages and three types of variants
324
M. Veigt et al.
22.5.2 System Modeling Information exchange is a prerequisite for distributed systems, consequently the information flow has to be modeled as a first step for the envisioned distributed control system. Based on a multi-agent platform every involved logistic object is represented by a software agent. Following this object driven approach we have to differentiate and classify different kinds of agents. Product-Agent:
The intelligent product plays the role of the decision maker. Especially this agent incorporates a high degree of autonomous control and decision making methods for choosing a type variant and to determine the manufacturing step order. To collect needed information the product agent has to send requests to other product agents, station agents and hardware agents. This will be a peer to peer communication horizontally at the same level. The information about current customer order comes vertically from the agent that has the context information coming from the user interface. Station-Agent:
The station agent comes with its own context information and the current changes such as machine failure. The main role of this agent is managing the order processing of registered/scheduled products for the station. Depending on the preconfigured abilities the station agent can also make decisions about the current capacity or a feasible step at the station and announces machine failure on request. The original context information is important for this agent and comes vertically at the beginning of the life time of this agent. Current changes such as new customer order will be uninteresting at operating time. Workshop-Agent:
This agent manages and shares environment information about the assembly system. It interacts also with the user configuration interface and represents some kind of common ERP systems. This agent is located on the higher level 4 (see Fig. 22.3) and dynamically induces the instantiation of other needed agents at MES level depending on the current demand. Product agents, for example, are created when a physical part enters the assembly-line. Mediating-Agent
The interfacing between the control system that is actually based on MAS and the physical components of the assembly-line hardware is also designed to be done by using agents. Similar to central control approaches these agents are triggered by real world signals (e.g. sensors) which then induce further activities. These agents act as links between the hardware (automation level) and the agents taking the decision and controlling the process. The described model is completely based on the multi-agent structure of Java Agent Development Toolkit JADE which is considered as the leading open-source
22
Autonomous Control in Production Planning and Control
325
ERP «interface» Information Exchange
JADE-Platform Workshop Agent 1
1 1 *
* Station Agent
Product Agent *
** *
* 1 1
*
1
Production Management and Process Control
1
RFID Management Agent
Plant Management Agent
1 «interface» Hardware Abstraction Layer (HAL)
«interface» Specific Plant Interface
Physical Structure Fig. 22.5 Interaction-model of the Agents
agent platform in academia. The complete interaction and communication model is depicted in Fig. 22.5.
22.5.3 Physical Structure The physical system consists of a monorail conveying system and RFID technology characterized by a low frequency. The conveying system works with self propelling shuttles with the capability to carry loads of up to 12 Kg (Fig. 22.6). The modular designed system offers the freedom of future extensions. The current set-up for the illustrated scenario allows the flexible change of planned routes by using the system integrated monorail-switches that offer multiple paths (Fig. 22.7). Thus it is possible to remain on the main line or to deviate to a bypass.
326
M. Veigt et al.
Fig. 22.6 Self-propelling shuttles
Fig. 22.7 The whole assembly line
The applied 125 kHz RFID technology was customized for our purposes to work with the casted RFID tags. Antennas and reading modules are usually not designed to be capable to read out RFID Tags that are coated by metal. Several effects can be observed when inserting RFID Tags into metal: shift of frequency, attenuation of the reader signal, to name a few. While casting the metal frame of the automotive
22
Autonomous Control in Production Planning and Control
327
tail-lights the RFID Tags are inserted and then act as intelligent products in our scenarios. The described physical objects are designed as autonomous resources for the implemented system. The required interacting between agents as decision makers and the physical world is deployed using a “Hardware Abstraction Layer” which is directly connected to the mediating agents. This concept allows a structured access to nearly any physical item considering the findings from the point of view of data integration. Thus the conditional ability to understand and process the data from data sources are complied to build local decision-making [27]. Considering aspects from the point of view of data integration in autonomous logistic networks [27, 28] there is freedom in terms of future extensions of the system.
22.5.4 Implemented Methods The autonomous and decision-making methods originating from the CRC 637 were implemented within the product agents. During manufacturing processes the products move along some kind of “Product type Corridor”. This means that the availability of currently possible type variants changes depending on the progress of production. Using this flexibility and considering several context parameters, the product agents are able to re-decide their targeting product variant themselves [29]. This method is called the autonomous product construction cycle for manufacturing systems [29]. When achieving a decision-result for a final type variant, the next possible assembly steps are identified. The final decision-making for the next step is done with a multi-criteria mathematical evaluation method which is based on the fuzzy hierarchical aggregation [30]. The all-up situation is analyzed considering several criteria, such as waiting time at every alternative station, material stocks of stations and current customer orders.
22.5.5 Results The presented demonstrator reflects the first steps towards the implementation of autonomous control methods in industrial environments. The demonstrator shows the capability and applicability of these methods via a given real complexity of synchronization between physical and virtual world. It becomes clear that an emergence arises out of the decentralized approach. This becomes evident when applying intended malfunctions or failures to the system. The products are able to react to the new situation without central re-planning. It becomes also obvious that the technology fundamentals for Intelligent Products do already exist. We can state qualitatively that an increased robustness can be observed. The achieved results are dependent on a new information flow structure, which could be combined with current production planning and control systems.
328
M. Veigt et al.
22.6 Conclusion This contribution shows the structural changes that arise when integrating autonomous control into existing production planning and control systems. Hereby, a vertical and a horizontal integration can be established. Vertically the management level and the process control level merge into one level. Horizontally autonomous control units substitute coexisting Manufacturing-Execution-Systems. These autonomous control units can be implemented by using the multi-agent technology. Software agents represent logistic objects (part, machine, means of transport, etc.) in a virtual model of logistic system and take the role of decision makers instead of physical objects. The distributed decision-making and control affect the capability of flexible system reconfiguration and of reacting to malfunctions in real time. A prototype points out how to implement autonomous control on the object level as well as the abundance of required information exchange between involved objects. This prototype follows the horizontal approach which substitutes several MES with autonomous units. Results show that an execution of autonomous control on the process control level has been achieved. The information exchange takes place horizontally on the same level. For instance, information on current machine load, identification or localization of parts is available. Indeed, the availability of this information is essential for the applicability of autonomous control. However, this availability cannot be ensured by existing productions systems. The quantity and the quality of sensors is currently a limiting factor. In summary, the availability of needed information can be understood as the limiting factor for the implementation of autonomous control in production planning and control from a technical perspective. Considerably more research is necessary to determine how autonomous control can perform effectively with uncertain information and how this information can be made certain and reliable.
References 1. Kletti J (2007) Manufacturing execution system. MES, Springer 2. Arnold D, Furmans K (2009) Materialfluss in Logistiksystemen, 6th edn. Springer 3. Dickersbach JT, Keller G (2010) Produktionsplanung und -steuerung mit SAP ERP, 3rd edn. Galileo Press 4. Busch A, Dangelmaier W (2004) Integriertes supply chain management, 2nd edn. Gabler 5. Hachtel G, Holzbaur U (2009) Management für Ingenieure – Technisches Management für Ingenieure in Produktion und Logistik. ViewegCTeubner 6. Kuhn A, Hellingrath B (2002) Supply chain management. Springer 7. Kilger C, Hermann R (2006) Integriertes informationsmanagement in intralogistik-systemen architektur und potenziale. In: Arnold (ed) Intralogistik. Springer, Berlin, pp 212–223 8. Finkenzeller K (2008) RFID-handbuch, 5th edn. Hanser 9. Dangelmaier W, Franke W (2006) RFID – Leitfaden für die Logistik. Gabler 10. Duffie N (1982) An approach to the design of distributed machinery control systems. IEEE Trans Ind Appl IA 18(4):435–442
22
Autonomous Control in Production Planning and Control
329
11. Shen W, Norrie DH (1999) Agent-based systems for intelligent manufacturing; a state-of-theart survey. Knowledge Inf Syst 129(1) 12. Paolucci M, Sacile R (2005) Agent based manufacturing and control systems. CRC, Boca Raton 13. Busse M, Wöstmann F-J, Müller T, Melz T, Spieß P (2006) Intelligente druckgussteile. Metall (11):738–741 14. Jrennings N (1996) Applying agent technology. Plenary presentation at PAAM’96, London, pp 22–24 15. Khedro T, Genesereth MR (1994) The federation architecture for interoperable agent-based concurrent engineering systems. Concurrent Eng 125(2) 16. FIPA (2002) Foundation for Intelligent Physical Agents. Fipa standard status specifications. Retrieved from: http://www.fipa.org/repository/standardspecs.html. 17. Searle JR (1969) Speech acts: an essay in the philosophy of language. Cambridge University Press 18. Parunak HVD (2000) A practitioner’s review of Industrial agent applications. Auton Agents Multi-Agent Syst 389(3) 19. Parunak HVD (1999) Industrial and practical applications of DAI. In: Weiss G (ed) Multi-agent systems: a modern approach to distributed artificial intelligence. MIT, Cambridge, MA 20. Wooldridge M, Jennings NR (1995) Intelligent agents: theory and practice. Knowledge Eng Rev 115(10) 21. Shen W (2002) Distributed manufacturing scheduling using intelligent agents. Intell Syst 17(88) 22. Shen W, Barthès JP (1996) Computer supported cooperative environments for engineering design: a survey. Tech. 96-122. Retrieved from: http://imsg.enme.ucalgary.ca/CIG/papers/ csceed.pdf. CNRS UMR Heudiasyc, Université de Technologie de Compiègne, France 23. Scholz-Reiter B, Windt K, Freitag M (2004) Autonomous logistic processes – new demands and first approaches. In: Monostori L (ed) Proceedings of the 37th CIRP international seminar on manufacturing systems. Hungarian Academy of Science, Budapest, Hungaria, pp 357–362 24. Scholz-Reiter B, Freitag M (2007) Autonomous processes in assembly systems. CIRP Ann 56(2):712–729 25. Teti R, Kumara SRT (1997) Intelligent computing methods for manufacturing systems. Ann CIRP 46(2):629–652 26. Ganji F, Morales Kluge E, Scholz-Reiter B (2010) Bringing agents into application: intelligent products in autonomous logistics. In: Schill K, Scholz-Reiter B, Frommberger L (eds) Artificial intelligence and losistics (AILog) – workshop at ECAI 2010 27. Hans C, Hribernik K, Thoben K-D (2008) An approach for the integration of data within complex logistics systems. In: Haasis H D, Kreowski H-J, Scholz-Reiter B (eds) Dynamics in logistics, first international conference, LDIC 2007. Springer, Heidelberg, pp 381–389 28. Hribernik K, Hans C, Thoben K-D (2009) The application of the epcglobal framework architecture to autonomous control in logistics. In: 2nd International conference on dynamics in logistics, Springer 29. Windt K, Jeken O (2009) Allocation flexibility – a new flexibility type as an enabler for autonomous control in production logistics. In: 42nd CIRP conference on manufacturing systems 30. Rekersbrink H, Ludwig B, Scholz-Reiter B (2007) Entscheidungen selbststeuernder logistischer Objekte. Ind Manag 23(4):25–30
Index
A-priori knowledge, 230, 231 Active communication, 209 BananaHop protocol, 209, 210 experimental data losses, 210–212 signal attenuation analysis, 212, 213 TelosB sensor, 209 Adaptive hop limit, 204 Advanced-planning and scheduling-systems (APS), 316 Agent-based control, 229 agent activity model, 234 atomic activity, 234, 235 deliberation and execution tasks, 234 eager learning algorithms, 236 model exploitation, 237 ordering constraints, 234, 235 planning and scheduling, 236 strategic knowledge management, 237 autonomous decision-making, 242 inference, 243 model applicability and quality, 242, 243 situation awareness, 244, 245 learning agent, 237 agent life cycle, 239, 240 architecture, 237, 238 critic, 237, 238 design choices, 239 eager learning, 240–242 problem generator, 238, 239 Agent-oriented knowledge management (AKM), 231 concrete decision situation, 232 knowledge management activities, 231, 232 personal knowledge management, 231 real-world constraints, 233 strategies and prioritization, 233
AKM, see Agent-oriented knowledge management ALEM, see Autonomous logistic engineering methodology Apparel logistics, 295 article bundles allocation, 295, 296 autonomous control, 296, 297 disposition and control, customer orders, 295 supply chain, 296 ARM XScale processor, 217 Artificial neural network (ANN) based plausibility checking, 220, 221 Auto-identification technologies, 292 Automobile terminal, see also Vehicle logistics average parking area utilization, 283, 284 due date reliability, 286, 287 priority list, 281 schematic representation, 278 travel times, 284, 285 treatment times, 279 basic and full flexibility scenario, 284, 285 increased load and full flexibility scenario, 285, 286 Autonomous assembly line, 322 achieved results, 327 assembly scenario, 323 implemented methods, 327 physical system, 325–327 system modeling, 324, 325 Autonomous control, see also Apparel logistics; Vehicle logistics agent based manufacturing management, 277 common scheduling heuristics, 277 decision making and decision execution, 276
M. Hülsmann et al. (eds.), Autonomous Cooperation and Control in Logistics, c Springer-Verlag Berlin Heidelberg 2011 DOI 10.1007/978-3-642-19469-6,
331
332 dynamic behavior and production planning, 275 limitations, 286 cause and effect relationships, 298, 306 economic limitations, 305, 306, 308, 309 methodical limitations, 299, 300 organizational limitations, 302–304 practical scenarios, 306, 307 technical limitations, 300–302, 307 logistic target achievement, 277 organizational demands, 276 process-related demands, 276 simulation, 277 control method selection criteria, 279, 280 data evaluation scheme, 281, 282 due date reliability, 286, 287 limitations, 287, 288 model description, 277–279 model parameters, 282, 283 opportunities, 286, 287 parameter depth, 280 priority lists, 281 random method, 281 standard method, 281 travel time, 284, 285 treatment time, 284–286 utilization, 283, 284 Technical demands, 276 Autonomous control method, production networks, 32 due date (DUE) policy, 33 HBA, 33 pheromone based policy, 33 QLE policy, 32 Autonomous cooperating logistics processes, see also Data integration decision making, 248 description, 249 intelligent logistics objects, 247 IT logistics landscape, 247, 248 data source, 250, 251 EDIFACT EANCOM and SAP RFC, 250 intelligent logistics objects integration, 250, 251 sensor and sensor network integration, 251, 252 Logistics objects, 250 Autonomous logistic engineering methodology (ALEM) bottom-up modeling approach, 150 domain orientation, 151
Index flexibility and robustness, 149 model-usage orientation, 151, 152 nondeterministic system, 150 procedure model, 154, 155 shop-floor manufacturing system, 149 user orientation, 151 view concept, 152–154 Autonomous logistics processes, 77 autonomous control, 77 interaction effort for resource utilisation, 86, 87 interaction effort for team formation, 81, 87, 88 cooperative problem solving, 78, 79 decision-making, 77 effective resource utilisation, 81 emergent interaction pattern, 82, 83 requirements and related work, 81, 82 efficient resource utilisation, 83 requirements and related work, 83, 84 team formation interaction protocol, 84, 85 insufficient individual capability, 79, 80 investigated process, 78 problem decomposition, 88 resource utilisation efficiency, 80, 81 supply network management, 77 team formation mechanism, 89 Autonomously controlled production systems, 131 decision making capability, 131 evaluation and result, 138 dynamic production environment, 144 genetic algorithm (GAL), 139 harmonic material flow, 144 job related throughput time, 142 optimization, 143 problem instances, 138, 139 QLE method, 141, 142 realized processing time, 140 scheduling method, 140 flexible flow shop (FFS) problem, 132–134 heuristics scheduling, 134, 135 intelligent logistic object, 131 manufacturing, 135 decentralized decision-making, 135, 136 local information method, 136 self-organizing behavior, 135 retooling sequence, 145 shop-floor environment, 145 structural complexity, 132
Index BananaHop protocol, 210 data transfer rate, 210 energy consumption, 210 experiments, 212 radio link quality and the performance, 210, 211 signal attenuation analysis, 212, 213 temperature and humidity measurements, 209, 210 BLG LOGISTICS GROUP, 277 Business process modeling, 149 ALEM, 150 bottom-up modeling approach, 150 domain orientation, 151 flexibility and robustness, 149 model-usage orientation, 151, 152 nondeterministic system, 150 procedure model, 154, 155 shop-floor manufacturing system, 149 user orientation, 151 view concept, 152–154 complexity induced challenge, 157 local process modeling, 158, 159 model creation, 159 model modification, 159, 160 model testing, 160–162 decentralized autonomous system, 166 flexible flow-job system, 155, 156 information management strategy, 166 organizational independence, 162 apparel supply chain, 157, 162 collaborative complexity, 163 collaborative modeling, 164, 165 organizational structure, 162 problem areas, 163, 164
Capacity control system, 52 bid-price control (BP), 55 process_by_quota procedure, 53 Pseudo-code, 53 quota-based control (QUOTA), 55 resource allocation, 53–55 Cause and effect relationships, 298 Chemotaxis (CHE) policy, 33 Collaborative Research Centre 637 (CRC 637), 291 Communication volume, 201, 202 40-vertex topology, 202, 203 communication traffic, 198 Communities of autonomous units, 119 basic actions, 116 computational power and limitation, 123 bandactions, 126, 127
333 beyond computability, 124 Com (TM), 127 graph band, 125, 127 logistic application, 123 polynomial community, 127, 128 state transition relation, 125 control conditions, 118 environment, 115, 116 global control condition, 119 graph class expression and goal, 116 graph transformation rule, 114 graph- and rule-based framework, 113 job-shop scheduling problem, 120 autonomous and interactive behavior, 120 environment specification unit, 121 finite machine set, 120 select/process rule, 121 switch/remove rule, 122 multi-agent system, 113 parallel actions, 118 production network, 114 Complexity induced challenge supply chain scenario, 156, 157 Customer value creation, 20 autonomous cooperation-based technology, 18 customer service, 19 heterarchical logistics system, 18 neural network computational model, 20 self-organization, 18 supply chain management, 19 benefits and sacrifices, 15 customer-supplier emotional bond, 15 decentralized decision-making, 16 intelligent container, 20 financial cost, 20–22 non-financial cost, 20–22 psychic value, 20, 21 utility value, 20, 21 logistics business model, 16 customer demand, 16 dimensions, 17, 18 firm-specific capability, 17 marketing, 16 value exchange model, 17 perishable goods, 23
Data integration, 259 autonomous cooperating logistics processes, 259 data source scalability, 262, 263
334 decision system dynamism, 262 evaluation, 263 flexibility criteria, 261, 262 querying application, 263 reactivity and data volume scalability, 264 reliability criteria, 259–261 classification, 254 data warehouse, 254, 255 enterprise service bus, 256, 257 federated database-systems, 255 message-oriented middleware, 256 object-oriented databases, 257 operational data store, 255 peer-data-management systems, 257 semantic mediator, 258 service oriented architecture, 256 loosely coupled integration, 253 object-oriented integration, 253 tightly-coupled integration, 252, 253 Data warehouse, 254, 255, 262 Decentralized decision-making, 3 Decentralized routing method, 196 Deterministic linear programs (DLP), 47 Distributed logistic routing protocol (DLRP), 195 40-vertex topology, 203 adaptive hop limit, 204 communication volume, 204, 205 fixed, scenario-wide hop limits, 203, 204 highway network, 202, 203 route request forwarding, 203 architecture, 197 communication volumes, 201, 202 constraint, 195 costs and resource consumption, 205 decentralized routing method, 196 goods, vertices and vehicles interaction, 196, 197 message sizes, 198 assumption, 198, 199 communication traffic, 198 route announcement messages, 200, 201 route disannouncement messages, 201 route reply messages, 200 route request messages, 199, 200 multi-criteria routing, 197 route discovery messaging, 196, 197 scenario topologies, 196 DLRP, see Distributed logistic routing protocol
Index Eager learning, 240 vs. lazy learning, 236 description, 240–242 Economic limitations, 305, 306, 309 Electronic product code (EPC), 318 Embedded intelligent objects, food logistics, see Food logistics Employee experience, 291, 292 Enterprise service bus (ESB), 256, 257, 263 Enterprise-resource-planning-systems (ERP), 316
Federated database-systems (FDBS), 255, 262, 263 Fixed, scenario-wide hop limits, 203, 204 Flexible flow shop (FFS) problem, 132–134 Flexible flow shop scheduling, 98 decentralized decision-making, 98 honey bee concept, 100 pheromone-based strategy, 99, 100 production resource, 98 queue length estimator (QLE), 99 rational strategy, 99 scheduling and buffer clearance, 101 Flow-line manufacturing system, 101 first in-first out (FIFO), 102 hybrid model, 103 production line, 101 shop floor scenario, 101 Vensim DSS software, 102, 103 Food logistics, 207 active communication, 209 BananaHop protocol, 209, 210 experimental data losses, 210–212 signal attenuation analysis, 212, 213 TelosB sensor, 209 automated sensor data processing, 225, 226 communication protocols, 208, 209 communication range, active wireless sensors, 208 energy efficiency, decision algorithms, 208 Kriging method, 209 logistical planning process, 207 passive RFID communication, 213 vs. active communication, 213 data transfer rate, 215, 216 integrated systems, 216 RFID tags, 214 time window for identification, 215 required CPU time measurement, 221 calculation vs. communication energy, 223
Index calculation vs. communication energy, 224, 225 decision algorithms comparison, 221, 222 energy consumers, 223 required processor energy, decision making, 216 algorithms, 216, 217 artificial neural network, plausibility test, 220, 221 autonomous plausibility checking, 219, 220 dynamic combination of algorithms, 221 energy properties, applied wireless sensor nodes, 217 joule per decision, 216 spatial interpolation, Kriging method, 218, 219 temperature effect on shelf life, 217, 218 temperature prediction, 218 temperature monitoring, 207
Generic algorithms, 3 Groupage systems, 61 collaboration, 62 characteristics and advantages, 66, 67 joint operational planning, 64 levels, 70 multi-depot vehicle routing problem, 65 operational management, 63 prerequisites and obstacles, 67, 68 profit sharing scheme, 63 shipment assignment, 65 shipper collaboration problem, 64 exchange mechanism, 73 limits of cooperation, 69 business connection, 69 harmonized strategy and behavior, 68 profit-sharing phase, 69 vehicle scheduling, 70 motivation, 61 profit-sharing model, 73
Honey bee algorithm (HBA), 33
Information and communication level, 2 Information and communication technology, 6 Information systems, 316 functions and limitations, 316
335 automation, 317 ERP and APS, 316 MES, 316, 317 structure, 315 Intelligent container, 271
Job-shop manufacturing, 177, 178 Job-shop scheduling problem, 120 autonomous and interactive behavior, 120 environment specification unit, 121 finite machine set, 120 select/process rule, 121 switch/remove rule, 122
Knowledge management, 230 agent activities and temporal constraints, 233 autonomous decision-making, 242–245 compilation and maintenance, 237–242 minimal model, 234–237 agent-oriented knowledge management, 231 concrete decision situation, 232 knowledge management activities, 231, 232 personal knowledge management, 231 real-world constraints, 233 strategies and prioritization, 233 knowledge forms, 230, 231
Logistics business model, 16 customer demand, 16 dimensions, 17, 18 firm-specific capability, 17 marketing, 16 value exchange model, 17
Manufacturing flexibility, 174, 175 Manufacturing system, autonomous control, 135 decentralized decision-making, 135 information discovery method, 136–138 local information method, 136 self-organizing behavior, 135 Manufacturing-execution-systems (MES), 316, 321, 322 Material flow level, 2 Message-oriented middleware (MOM), 256 Methodical contributions and limitations, logistics, 93
336 computational complexity, 94 dynamic logistic environment, 93 generic algorithm, 93 information technology, 94 manufacturing process, 95 system complexity and dynamics, 95 Methodical limitations, 299 applicability and effectiveness, autonomous control, 299 decision making, 299 decision rules modification, 300 implicit knowledge, vehicle logistics, 299 slack time dispatch rule, 300 Multi-agent systems (MAS), see also Knowledge management, see also Knowledge management Multi-criteria context-based decision (MCCD) function, 198 Multiple autonomous control strategies, 97 decentralized system, 97 design of autonomous service rules, 103 first in-first out (FIFO), 103 pheromone-based, 103–105 queue length estimator (QLE), 105 result evaluation, 105, 106 flexible flow shop scheduling, 98 decentralized decision-making, 98 honey bee concept, 100 pheromone-based strategy, 99, 100 production resource, 98 queue length estimator (QLE), 99 rational strategy, 99 scheduling and buffer clearance, 101 generic matrix model, 109 machine breakdown or rush order, 110 matrix-like flow-line manufacturing system, 101 first in-first out (FIFO), 102 hybrid model, 103 production line, 101 shop floor scenario, 101 Vensim DSS software, 102, 103 micro-macro-link, 97 pheromone correction term, 107–109 pheromone-based scheduling, 98 self-organization, 98 simple non-autonomous service rule, 106, 107
Network capacity control in road haulage, 45 capacity control approach, 46 capacity control system, 52 bid-price control (BP), 55
Index process_by_quota procedure, 53 Pseudo-code, 53 quota-based control (QUOTA), 55 resource allocation, 53–55 computational experiment, 55 bid-price approach, 57 evaluation metrics, 56 gained revenue, 56 revenues percentage, 56, 57 customer demand, 45 decision making, 45 decision situation, 46 bid-price capacity control strategy, 52 dynamic decision problem, 47–49 literature, 46, 47 model-based determination of quotes and bid-prices, 49, 50 non-homogeneous Poisson process, 51 revenue management, 51 sales and capacity management, 58
Object-oriented databases, 257, 264 Operational data store (ODS), 255, 262 Optimal strategy of production network, 30 distributional coefficient, 30 material storage, 30 mixed controls, 32 piecewise constant function, 31 Pontryagin maximum principle, 31 Organization and management level, 2 Organizational contributions and limitations, 4, 5 functional perspective, 11 institutional perspective, 11 instrumental view, 12 optimal control theory, 13 organizational measure, 13 process-related perspective, 12 self-organization, 11 team formation, 14 Organizational demands, 276 Organizational independence, 157 apparel supply chain, 157, 162 collaborative complexity, 163 collaborative modeling, 164, 165 organizational structure, 162 problem areas, 163, 164 Organizational limitations, 302 organisational culture, 302, 303, 308, 309 resistance against changes, 304 Organizational perspective, 3, 4
Index Passive RFID communication, 213 data transfer rate, 215, 216 integrated systems, 216 RFID tags, 214 time window for identification, 215 vs. active communication, 213 Peer-data-management systems, 257 Pheromone based policy, 33 Plausibility checking, 219 ANN based, 220, 221 autonomous, 219, 220 dynamic combination of algorithms, 221 Kriging, 225 PPC, see Production planning and control Practical contributions and limitations, 271–273 Process-related demands, 276 Production networks, 27 cross-company owned network, 27 input-to-state stability concept, 28 limitations of autonomous control, 40, 41 modeling and control, 29 autonomous control method, 32–34 central planning method, 30 distribution coefficient, 30 manufacture process, 29 network property, 34, 35 optimal strategy, 30–32 supply network, 29 production planning and control (PPC), 27 single production system, 28 stability, 35 gain-matrix, 37 Lipschitz continuous function, 35 local small gain condition (LSGC), 37 locally input-to-state stable (LISS) system, 36 Lyapunov function, 36 ordinary differential equation (ODE), 35 parameter constellation, 38, 39 possible generalization, 39, 40 worst-case approach, 38 Production planning and control (PPC), 313 automation pyramid, 314, 315 autonomous assembly line, 322 achieved results, 327 assembly scenario, 323 implemented methods, 327 physical system, 325–327 system modeling, 324, 325 autonomous control, 320 integration, 320, 321
337 manufacturing execution system, 321, 322 central planning and control, 313 challenge, 313 information systems, 315 automation, 317 ERP and APS, 316 MES, 316, 317 technical approach, 317 agent based control systems, 318 autonomous methods and decision algorithms, 319, 320 unique identification and localization, 318 Production processes, autonomous control, 169 basic principle, 173, 174 industrial practice, 170 allocation change of production items, 171, 172 industrial management, 170 order sequencing, 171 product variety, 171 sheet metal production, 170 steel industry, 172 manufacturing flexibility, 174 methods, 176 ant pheromone, 179 centralized production planning and control, 180 CRC 637, 178 generic algorithm, 178 job-shop manufacturing, 177, 178 material flow system, 177 rational strategy, 179 technological capability, 176 order allocation flexibility, 169, 175 order allocation practice, steel making, 181 dynamic order allocation practice, 183, 184 steel industry phenomenon, 181, 182 order management, 169 research, 185 Programmable logic controller (PLC), 317
Queue length estimator (QLE) policy, 32
Resource utilisation, 80 effective, 81 emergent interaction pattern, 82, 83 requirements and related work, 82 efficient, 83
338 lot sizing, 80 requirements and related work, 83, 84 service provider, 80, 81 team formation interaction protocol, 84, 85 RFID-interrogator, 297 RFID-transponder, 294, 317 Route discovery messaging, 196, 197 route announcement messages, 200, 201 route disannouncement messages, 201 route reply messages, 200 route request messages, 199, 200
Self-organization principle, 5 Semantic mediator, 258, 264 Service oriented architecture (SOA), 256, 263 Software bugs, 291, 292 Steel making, 183 dynamic order allocation practice, 183, 184 steel industry phenomenon, 181, 182 Supply chains, 156, 157
Technical demands, 276 Technical limitations, 300 barcode and RFID, 301, 302 computing power, 301 economic limitations, 302, 305, 306 equipment, 300, 301, 308 logistic objects positioning, 302
Index mobile communication, 302 operation time, computing system, 301 Technological contributions and limitations, 191 data integration mechanisms, 193 decentralized decision-making, 192 distributed logistic routing protocol (DLRP), 192 flexible order processing, 191 information and communication technology, 191, 192 knowledge management, 193 multi agent systems (MAS), 192 TelosB sensor active communication, 209, 210, 213, 217 ANN based plausibility checking, 220 required CPU time, 221 required processor energy, 217
Vehicle identification number (VIN), 293, 294 Vehicle logistics, see also Automobile terminal sensor system, 293 storage allocation, 294 vehicle identification, 293, 294 wearable computing system, 294, 295 Vehicle routing problem, 46 Virtual cars path, 278, 279
Wearable computing system, 294, 295
