Commitment and Evolution Connecting Emotion and Reason in Long-term Relationships István Back
ii
The research describ...
29 downloads
1146 Views
4MB Size
Report
This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form
Commitment and Evolution Connecting Emotion and Reason in Long-term Relationships István Back
ii
The research described in this thesis was carried out under the auspices of the Interuniversity Center for Social Science Theory and Methodology (ICS) and the Faculty of Social and Behavioral Sciences (GMW) at the University of Groningen (RuG). Funding was generously provided by the Ubbo Emmius Bursary (2003) and by the Netherlands Organization for Scientific Research (NWO). © 2007 by István Back All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means without written permission of the author. The document was typeset using LATEX 2ε and Stasinos Konstantopoulos’s RuGthesis.cls. Printed by Mesterprint Kft., Budapest, Hungary. ICS Dissertation series (nr. 133)
ISBN 978-90-367-3113-3
R IJKSUNIVERSITEIT G RONINGEN
Commitment and Evolution Connecting Emotion and Reason in Long-term Relationships
Proefschrift
ter verkrijging van het doctoraat in de Gedrags- en Maatschappijwetenschappen aan de Rijksuniversiteit Groningen op gezag van de Rector Magnificus, dr. F. Zwarts, in het openbaar te verdedigen op maandag 17 september 2007 om 14.45 uur
door
István Henrik Back geboren op 19 november 1979 te Boedapest, Hongarije
iv
Promotor:
Prof. dr. T.A.B. Snijders
Copromotores:
Dr. H. de Vos Dr. A. Flache
Beoordelingscommissie:
Prof. dr. M.W. Macy Prof. dr. S.M. Lindenberg Prof. dr. A. Riedl
Acknowledgements I would like to thank my supervisors, Henk de Vos, Tom Snijders, and especially Andreas Flache for their invaluable input throughout the last four years. Their encouragement, inspiring ideas and complementary expertise was instrumental to carrying out this interdisciplinary piece of research. I received further assistance from many others working in the graduate school ICS in Groningen, Utrecht and Nijmegen, most notably Vincent Buskens who always kept a watchful eye on my research, provided me with ideas and helped to carry out a large proportion of my experiments; Károly Takács and Michael Mäs, who were always ready to read and discuss my drafts; Rita Smaniotto who helped me keep my enthusiasm for evolutionary theory; Sigi Lindenberg; Jeroen Weesie; Inneke Maas; Frans Stokman; Richard Zijdeman, Eva Jaspers, Ellen Verbakel, Nienke Moor, Janneke Joly and Stefan Thau; Jessica Pass, Jacob Dijkstra, Christian Steglich, Lea Ellwardt and Jurre van den Berg. I thank Michael Macy and graduate students at the Sociology Department of Cornell University, especially Arnout van de Rijt and Ma Li for making my visit there so rewarding not only professionally but also personally; and David Sloan Wilson at Binghamton University for thought-provoking conversations about human evolution. I owe big thanks to Ji Wenxi (Wendy) who not only provided me with invaluable support during my experimental work in China but continues to be a window into the oriental mind and thinking; Fan Xuejuan at East China Normal University, Xu Bo and Xu Longshun at Fudan University who generously provided the means to carry out my experiments in Shanghai; Zhao Kanglian at Nanjing University. Xu Yu, Gerbren Kuiper and Huixin, Yorgos Vleioras, Justin Park, Simon Dalley, Bori Takács, Tamás Bíró, Ela Polek, Andrea Szentgyörgyi, Gábor Imre, Li Kun and Wang Zhuo for keeping me company in Groningen; and Evelien de Roos, the best land-lady in the Netherlands. Finally, I would like to thank my friends in Hungary, Miki Rosta, Levente Skultéti, Attila Máté, György Hermann, Zoli Gedei, Kristóf Bajnok and Laura Radics, for their friendship which has been one of the key sources of motivation behind this piece of work.
v
Contents 1
Introduction 1.0.1 A brief word on “commitment” . . . . . . . . . . . . . . . 1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1.1 Commitment to a course of action . . . . . . . . . . . . . 1.1.2 Interpersonal commitment . . . . . . . . . . . . . . . . . 1.2 Toward an evolutionary explanation . . . . . . . . . . . . . . . . 1.2.1 Separating ultimate and proximate explanations . . . . . 1.2.2 How evolutionary theory helps to explain seemingly irrational behavior . . . . . . . . . . . . . . . . . . . . . . . 1.2.3 Constructing an ultimate explanation for interpersonal commitment . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.4 Proximate mechanisms for interpersonal commitment . 1.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4 Outline of chapters . . . . . . . . . . . . . . . . . . . . . . . . . .
17 20 21 22
I
An ultimate explanation
25
2
The Competitive Advantage of Commitment 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . 2.2 Model . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.1 Modeling strategies . . . . . . . . . . . . . 2.2.2 Evolutionary dynamic . . . . . . . . . . . . 2.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.1 Simulation setup . . . . . . . . . . . . . . . 2.3.2 The unconditionality of Commitment . . . 2.3.3 Explanation: the importance of strong ties 2.3.4 Sensitivity to initial parameters . . . . . . . 2.4 Discussion and Conclusion . . . . . . . . . . . . .
27 28 32 33 37 38 39 41 46 48 50
vii
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
1 4 5 6 7 12 15 15
viii 3
4
II 5
CONTENTS The Evolutionary Advantage of Commitment 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.1 Modeling strategies . . . . . . . . . . . . . . . 3.2.2 Evolutionary dynamic . . . . . . . . . . . . . . 3.3 Conjectures . . . . . . . . . . . . . . . . . . . . . . . . 3.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4.1 Initial parameters . . . . . . . . . . . . . . . . . 3.4.2 Stability . . . . . . . . . . . . . . . . . . . . . . 3.4.3 The importance of interpersonal commitment 3.5 Discussion and conclusions . . . . . . . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
Fairness and Commitment under Inequality 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.1 Modeling strategies . . . . . . . . . . . . . . . . . . . . . 4.2.2 Evolutionary dynamic . . . . . . . . . . . . . . . . . . . . 4.3 Conjectures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.1 Initial parameters . . . . . . . . . . . . . . . . . . . . . . . 4.4.2 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.3 The importance of interpersonal commitment . . . . . . 4.4.4 The relative importance of fairness, commitment and capability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.5 Sensitivity to initial parameters . . . . . . . . . . . . . . . 4.5 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . .
Proximate explanations Commitment Bias 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3 Experimental design . . . . . . . . . . . . . . . . . . . . . . . . . 5.3.1 Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3.2 Sample . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4.1 Hypothesis 5.1 and 5.2 – The commitment bias . . . . . . 5.4.2 Hypothesis 5.3 – Effect of affect . . . . . . . . . . . . . . . 5.4.3 Hypothesis 5.4 and 5.5 – Cross-cultural similarities and differences . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4.4 Self-reported reasons for exit . . . . . . . . . . . . . . . . 5.5 Discussion and conclusions . . . . . . . . . . . . . . . . . . . . .
53 54 56 56 58 59 60 61 61 62 67 71 72 73 74 76 77 79 80 80 81 81 84 85
89 91 92 95 98 100 100 100 101 103 103 107 107
CONTENTS
ix
5.5.1 5.5.2
The evolutionary roots of commitment . . . . . . . . . . 109 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . 111
6
Commitment and Networking under Uncertainty 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2 Theory and Hypotheses . . . . . . . . . . . . . . . . . . . . . . . 6.3 Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3.1 Sample 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3.2 Sample 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3.3 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3.4 Manipulation of uncertainties . . . . . . . . . . . . . . . . 6.3.5 Manipulation of dilemma . . . . . . . . . . . . . . . . . . 6.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4.1 Final sample . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4.2 Hypothesis 6.1 – Social uncertainty and commitment . . 6.4.3 Hypothesis 6.2 – Resource uncertainty and commitment 6.4.4 Hypothesis 6.3/6.4 – Interaction between uncertainties . 6.4.5 Hypothesis 6.5/6.6 – Trust and Optimism . . . . . . . . . 6.5 Robustness of results across different dilemmas . . . . . . . . . 6.6 Conclusion and Discussion . . . . . . . . . . . . . . . . . . . . . 6.6.1 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . .
113 114 115 121 122 122 123 124 125 125 126 127 127 128 129 130 131 132
7
Conclusions 7.1 Summary of results . . . . . . . . . . . . . . . . . . . . . . . . . 7.2 General discussion . . . . . . . . . . . . . . . . . . . . . . . . . 7.2.1 In defense of evolutionary theory in the social sciences 7.2.2 Placing our work . . . . . . . . . . . . . . . . . . . . . . 7.2.3 Innovations of the present work . . . . . . . . . . . . . 7.2.4 Possible criticism . . . . . . . . . . . . . . . . . . . . . . 7.2.5 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . 7.2.6 Avenues for future research . . . . . . . . . . . . . . . .
133 134 136 137 138 141 142 143 144
. . . . . . . .
Bibliography
147
Summary
159
Samenvatting - Dutch summary
163
Összefoglalás - Hungarian summary
167
Zhai Yao - Chinese summary
171
A Analytical solution of the simplified dilemma
175
x
CONTENTS
B Parameter values used in ecological simulations
177
C Pseudocode of simulation core
179
D Parameter values used in evolutionary simulations
181
E Pseudocode of evolutionary dynamic
183
F Experiment instructions F.1 Initial instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . F.2 Instruction text from the experiment . . . . . . . . . . . . . . . . F.3 Screenshot from the experiment game . . . . . . . . . . . . . . .
185 185 185 187
G Experiment instructions 189 G.1 Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189 G.2 Screen shots from the experiment . . . . . . . . . . . . . . . . . . 190
Chapter 1
Introduction “We may call the part of the soul whereby it reflects, rational; and the other with which it feels hunger and thirst and is distracted by sexual passion and all the other desires, we will call irrational appetite, associated with pleasure in the replenishment of certain wants... What of that passionate element which makes us feel angry and indignant? Is that a third, or identical in nature with one of those two?” —Plato, The Republic
The tendency to establish lasting personal relationships is a fundamental aspect of human sociality. Throughout life we build friendships, collect acquaintances, forge business alliances, become attached to intimate partners. Many of these relationships follow us through our lives and integrate us into a complex social fabric of interpersonal connections. At the same time, establishing and maintaining long-term relationships involves substantial investment of one’s time, effort and other resources. Moreover, many relationships by definition require exclusivity. For example, we can only have one best friend at a time, in many cultures only one spouse, and in many business settings only one supplier of some product. To a certain extent all relationships, i.e. non-exclusive ones as well, are competitive with each other, given that we have finite attention and resources. This means that we occasionally have to forgo relationships with potentially better alternative partners. And to complicate matters, even when we do our best to invest in a relationship, we have to live with the risk of being dumped for someone else or unknowingly being taken advantage of by our partner. Why do people establish and maintain long-term relationships when these are costly, risky and exclusive? A simple but powerful answer from rational 1
2
Chapter 1. Introduction
choice theory is that it is in their best interest to do so. More precisely, people become committed to each other if and only if the benefits of having a relationship outweigh its maintenance costs and its alternative costs. In particular, having a long-term relationship with a partner provides valuable information about the trustworthiness of the partner compared to other partners (trust explanation, see Kollock, 1994; Yamagishi and Yamagishi, 1994; Yamagishi et al., 1994) and at the same time creates a strategic incentive to cooperate in order to avoid retaliation and stabilize long-term mutual collaboration (reciprocity explanation, see Trivers, 1971; Friedman, 1971; Axelrod, 1984; Fehr and Schmidt, 1999; Fehr and Gächter, 2002; Falk et al., 2001). But at the same time, there seems to be much more to long-term interpersonal relationships than just trust and reciprocity. There are numerous cases, for example, when people keep relationships even after their partner has proved to be untrustworthy (e.g. Roy, 1977; Strube, 1988; Rusbult and Martz, 1995). There are also examples of relationships where a partner has no means of reciprocating in the future (e.g. Monahan and Hooker, 1997). What is it that makes battered wives return to their abusive husband when there are hardly any prospects for change? And why does someone take care of a lifelong partner with Alzheimer’s disease who will never be able to recognize the caretaker? Why do subjects in controlled laboratory experiments give costly gifts to their long-term exchange partners when their identity will never be revealed to each other? A great wealth of empirical evidence suggests that people are engaged in long-term relationships with their full emotional repertoire (cf. Baumeister and Leary, 1995). People create social relationships with great ease even in the absence of materialistic benefits or other ulterior motives, and strongly resist the dissolution of these relationships, well beyond rational considerations of practical advantages. Many of the strongest emotions people experience in their life, both positive and negative, are linked to long-term relationships. The evidence suggests that being accepted, included, or welcomed leads to positive emotions such as happiness, elation, contentment, and calm, whereas being rejected, excluded, or ignored leads to anxiety, depression, grief, jealousy, and loneliness, etc. Indeed, the evidence is sufficiently broad and consistent to suggest that one of the basic functions of emotion is to regulate behavior so as to form and maintain social bonds (Baumeister and Leary, 1995). There is further evidence that people observe and evaluate alternative partners with a biased vision, systematically dependent on how committed their current relationship is (Johnson and Rusbult, 1989). Moreover, we know that even in anonymous exchange settings, positive emotions develop toward frequent exchange partners, and toward the relationship itself, being perceived as an object of value (Lawler and Yoon, 1996). These emotions provide a positive feedback for commitment behavior and lead to a systematic divergence from instrumental rationality.
3 But why is it that our relationship-related emotions are so often out of tune with what is usually regarded as rational? What is the source of emotions that make us consistently more committed than our best interest seems to dictate? Is there, in fact, something fundamentally rational behind seemingly irrational commitments? In order to resolve the paradox between rational and emotional explanations of interpersonal commitment, we put forward an evolutionary explanation. During countless years of prehistoric evolutionary adaptation in the human ancestral environment, people lived together in small groups and fought for daily survival in a world more hostile than today’s (Sterelny, 2003). With many of the formal and informal helping institutions of modern society missing, people had to rely on interpersonal relationships to a much larger extent than today. Sometime during the Pleistocene epoch (roughly 1.8 million years to 12 thousand years before the present) humans moved from rain forests to the savannah, which increased the need for collective hunting and mutual protection from large predators. This in turn created a selection pressure for increased social complexity (c.f. Smaniotto, 2004). At the same time, life-threatening situations produced more opportunities for bonding and deep friendships. Being capable and willing to establish and maintain long-term stable relationships substantially increased one’s survival and reproductive chances. As a consequence, those whose cognitive arsenal was equipped with better tools and stronger preferences for making interpersonal commitments gradually increased their presence in the population over many generations (cf. Nesse, 2001a). In lack of a direct test, evolutionary theories are difficult to empirically falsify and therefore problematic to find convincing support for. Therefore, our strategy in this dissertation is twofold. We first examine a theory of natural selection acting on commitment in closer detail in Part I (An ultimate explanation)1 . The main motivating question for this effort is: Could a trait of interpersonal commitment have been selected for in human evolutionary history, especially in the face of other, more or less cooperative, traits? Building on previous work (especially de Vos et al., 2001) that relies on anthropological knowledge about conditions of the human ancestral environment, we create formal computational models of the ancestral environment. The purpose of these models is to test the internal consistency of an evolutionary theory about deeply rooted (or “hardwired”) emotions that facilitate interpersonal commitments. Then, in Part II (Proximate explanations) we move on to empirically test the existence of an evolved commitment trait. The main question this part addresses is: Are there features of contemporary social behavior that are in line with an ancestral trait for commitment but cannot readily be explained by 1 See
more about the important distinction between ultimate and proximate explanations in evolutionary theory under Section 1.2.1 of this chapter.
4
Chapter 1. Introduction
simpler, existing theories? In order to test the existence of such a proximate mechanism (which we term the commitment bias), we conducted laboratory experiments at six locations in three different countries (the Netherlands, USA and China). In particular, we aimed to find support for mechanisms that are difficult to reconcile with current exchange theoretical and (social) psychological theories but become intelligible in light of the evolutionary explanation. In the remainder of this introductory chapter we are first going to clarify an important issue about the use of the word commitment. We then address the vast literature of commitment from philosophy, economics, game theory, exchange theory, psychology, sociology, and evolutionary psychology, pointing to how the dichotomy of emotional and rational explanations permeates the subject throughout these disciplines. Building on this broad background, we set out to construct an evolutionary theory that aims to bring the two sides of the dichotomy closer to each other and thus present the diverse literature of commitment in a new light.
1.0.1
A brief word on “commitment”
Before we turn to the substantial discussion of commitments, a brief clarification of the term itself is inevitable. The word “commitment” is used excessively in different meanings, within different contexts, which may lead to misunderstandings. It also creates seemingly unrelated research lines in various disciplines across the humanities and behavioral sciences. The first known record of the word entering the English language is from 1386, when Geoffrey Chaucer advised “commit the keeping of your person to your true friends..., who are the best «physicians» and most reliable help and healing” (Wyatt, 1999). Thus, in its original sense, commitment is a promise or threat, pledge, agreement, contract or dedication, made to oneself or to others, to do something or to act in a certain way in the future. “Being committed to protect one’s country from enemies” or “committing oneself to not getting married” are examples. Commitment in these cases is similar in meaning to persistence or consistence (see Section 1.1.1 “Commitment to a course of action” below). By extension of meaning, commitment came to refer to a bond, or loyalty toward a social entity, such as an organization, a group of people, or another person. The basis for this extension is that in such cases one acts in accordance with one’s expressed or understood promise to the entity, and membership therein. A friendship, a marriage vow, an employment contract, or simply refraining from extra-couple romance are examples. Commitment in this sense is related to meanings of belonging, stay behavior, loyalty or faithfulness (see Section 1.1.2 “Interpersonal Commitment” below). Arguably, these different meanings are not independent, and a closer look reveals a number of common characteristics. Firstly, commitment always re-
1.1. Background
5
quires behavioral consistency, in other words acting repeatedly in the same way with regard to the target of commitment. Secondly, commitment entails opportunity costs for the individual due to sacrificing potential rewards from alternative courses of action, that are not explored due to behavioral consistency. Finally, commitment is always temporally embedded – it has a duration in time, or at least it is in some sense about the future. It is by definition continuous in time because one cannot uphold the same commitment in disjoint fractions of time. Given these conceptual similarities behind different forms of commitment, it is surprising that hardly any interdisciplinary research has systematically explored links between commitment in the action and in the interpersonal sense. This is not our major undertaking either but as we will demonstrate through a brief literature review below, there is at least one crucial point on which most theories of commitment converge. This common point is the duopoly of two competitive explanations: one that advocates rational reasons and another that points to deeply rooted emotions. Ignoring either type of explanation leaves a theory potentially vulnerable to criticism by the other side. Our goal is therefore to derive and test hypotheses within an evolutionary framework that is able to accommodate both types of explanations and resolve some of the contradictions arising between them within the context of interpersonal commitment.
1.1
Background
The idea of interpersonal commitment is conceptually embedded into the more general notion of commitment to a course of action. To identify the implications of the more general concept for the more specific, we start our theoretical discussion with this broader idea of commitment. From here we proceed to our core topic of interpersonal commitment, the tendency to maintain long-term relationships. The existing literature of interpersonal commitment can be separated into two, largely disconnected fields. The first field, researched mostly by economists, focuses on exchange and social networks. The second, researched mostly by psychologists, is more directed at close relationships, such as married and romantic couples. We review some of the most important contributions within each field, pointing to the presence of the emotion-rationality dichotomy throughout. Finally, we briefly touch upon a hybrid area, organizational commitment, which is closest to business and management research, although it originally grew out of the psychological field of interpersonal commitment.
6
1.1.1
Chapter 1. Introduction
Commitment to a course of action
Committing ourselves to a course of action means that we voluntarily give up some of our freedom of choice, by agreeing to do (or not to do) something at some point in the future. The willingness to make such commitments has long intrigued scientists and philosophers alike, going back as far as ancient times. The paradoxical benefits of this seemingly self-defeating behavior was already recognized by ancient Greeks. Xenophon, a talented general, when facing a superior enemy, ordered his troops to take up a position with their backs to an impassable ravine (cf. Schelling, 2006) in order to eliminate all their routes of escape. By doing so he signaled both to the enemy and to his own men that there was no alternative for survival, except victory. A major theoretical advance came in 1785 with Immanuel Kant’s “Grundlegung zur Metaphysik der Sitten”, where he proposed to distinguish between two sources of commitment (cf. Levinger, 1999). He argued that commitment, on the one hand, can grow out of desire or affection. In the case of commitment of desire, people act out of inclination, for example, because they like to or enjoy it. Kant considered this type of commitment transient and therefore weak and untrustworthy. The other form of commitment stems from duty or moral obligation, in which case people act in accordance with principles. Kant argued that this type of commitment is more enduring and far better morally. With this theoretical distinction between “having to” and “wanting to”, Kant essentially created the fundamental dichotomy between rational and emotional explanations that still dominates the discourse over commitment. The next major contributor to the theory of commitment was Thomas Schelling with his seminal book “The Strategy of Conflict” (1963). For Schelling, commitment is a strategic tool, deliberate action, the purpose of which is to influence someone else’s choices. Schelling recognized the importance of being able to make commitments in situations where each actor’s outcome mutually depends on other actors’ actions. In such situations each actor needs to take into account what others are likely to do next. The fact that one makes a commitment to act in a certain way radically alters the expectations and decision processes of others (Schelling, 2006). The very possibility to make commitments is a key mechanism for achieving collectively desirable outcomes that are otherwise difficult to agree on (see e.g. Raub, 2004). According to Becker’s side-bet theory (1960), making a commitment links investment in an extraneous interest (a side-bet) with a consistent line of action. In Becker’s example, a man wants to buy a house. The man makes an initial offer of sixteen thousand dollars to the owner. The owner insists on having twenty thousand. Our well-prepared buyer, however, reaches into his pocket to produce certified proof that he had made a bet of five thousand dollars with a third-party that he will not pay more than sixteen thousand for the house. The seller has no choice but to accept the buyer’s standpoint.
1.1. Background
7
In this example, the buyer uses a “credible threat” (a commitment) to modify his own payoff structure, thus also modifying the strategic interdependence between the two. Becker’s theory has received extensive attention, and was widely tested empirically, albeit with mixed success (cf. Cohen and Lowenberg, 1990; Wallace, 1997). Another conceptualization of side-bets is voluntary hostage posting (e.g. Raub, 2004). Hostage posting means surrendering an object of value to a trustor in order to increase trust in the trustee’s willingness to uphold the promise (commitment) made to the trustor. The hostage promotes trust (at least in the economic sense) by binding the trustee through reducing his incentives for abusing trust, by providing compensation for the trustor in case trust is abused, and by serving as a signal for the trustor about unobservable characteristics of the trustee that are related to the trustee’s opportunities and incentives for abusing trust (Raub, 2004; Snijders and Buskens, 2001). An interesting case of commitment is the tendency to escalate investment in a failing course of action, in other words, “throwing good money after bad” (cf. Karlsson et al., 2005; Brockner, 1992; Staw, 1976, 1997). Also known as the sunk cost effect, this motivates people to continue investment in a project despite unsuccessful prior investments of money, effort, or time (Arkes and Blumer, 1985). It is important to recognize that thinking in terms of sunk costs is a departure from rational calculation, in the sense that it distorts actual costs and benefits associated with possible outcomes.
1.1.2
Interpersonal commitment
Interpersonal commitment, or becoming committed to long-term partners, is regarded as a special case of commitment to a course of action by some game theorists, economists and also others (cf. Frank, 1988; Nesse, 2001a). The core idea behind this association is that commitment in long-term relationships is based on an implicit or explicit promise to stay with the partner, and to uphold a general conduct that is aligned with the interests and expectations of the partner. This dissertation focuses on long-term interpersonal relationships, such as marriage, friendship and acquaintanceship, asking the question: why do people become committed to each other when it is seemingly not in their best interest? Just as commitment to uphold a certain course of action entails a seemingly irrational decision to reduce one’s set of available choices in the future, interpersonal commitment involves sacrificing interaction with potentially superior, alternative partners. Yet, as the former type of commitment proves to be not only rational but, in fact, essential for success in society, could the same be said about interpersonal commitment? The tentative answer is yes.
8
Chapter 1. Introduction
Commitment, exchange and uncertainty In social exchange, two or more actors exchange some form of material or social benefit among each other in order to arrive at an advantageous outcome. Exchange theory presumes that people exchange repeatedly with the same actors when success occurs but move to others when failure occurs. The underlying mechanism may be simple reinforcement learning (Homans, 1961; Emerson, 1972; Macy and Flache, 2002) or rational choice (Kollock, 1994; Cook and Whitmeyer, 1992). Exchange often motivates actors to unilaterally modify the balance of exchange to their own advantage without prior knowledge of the partner, in other words to cheat them in some way. This inevitably leads to uncertainty about the outcome of the exchange for both partners. When actors repeatedly exchange resources, they learn more about one another, find each other more predictable, develop mutual trust, and infer that they have similar orientations to the exchange task (Lawler, 2001). Therefore, a standard insight of exchange theory is that frequent exchange with the same partner reduces uncertainty about cheating, and thus decreases the likelihood of exchanging with strangers. More specifically, the uncertainty-reduction hypothesis was tested by Kollock (1994) who showed that commitment is more likely to form in markets where the quality of the products is unobservable at the time of the exchange. Kollock (1994) also simulated different market environments under controlled laboratory conditions. In one condition (high uncertainty), sellers could deceive their potential buyers about the quality of the product they were selling. In the other condition (low uncertainty), it was not possible to deceive buyers. A key finding of Kollock’s experiment was that commitment formation between a particular seller and a particular buyer occurs more frequently in the high-uncertainty condition than in the low-uncertainty condition. In the same vein, Yamagishi and Yamagishi (1994) argue that committed relations give a solution to the problem of uncertainty, for multiple reasons. First, committed partners accumulate information about each other over time. Second, mutually committed people enact “hostage-taking” behaviors (Raub, 2004) – ranging from the formation of mutual emotional attachments to the establishment of relation-specific assets (Helper and Levine, 1992). Hostage-taking behaviors provide deterrence against unilateral defection (Shapiro et al., 1992). Finally, conditionally cooperative strategies such as Tit-for-Tat can be used to control each other’s behavior (Axelrod, 1984). The main underlying argument for the uncertainty hypothesis is that individuals tend to avoid unpredictable or uncertain decision contexts (Tversky and Kahnemann, 1974; Kahneman and Tversky, 1979, 1996), which are created by a lack of first-hand knowledge about a potential partner’s trustworthiness (“social uncertainty”). But is the trust problem the only source of
1.1. Background
9
uncertainty in social exchange? Different exchange partners have different resources and may offer different benefits. The size and range of these potential benefits leads to a conceptually new source of uncertainty. Does this kind of “resource uncertainty” also increase commitment, independently from social uncertainty? We will examine the question of resource-inequality between exchange partners more closely in Chapter 4, and return to the concept of resource uncertainty in Chapter 6. Yamagishi and Yamagishi (1994) list several reasons for the difficulty people have in leaving a committed relationship even when it becomes a liability. One is that the mutual attraction and loyalty that have developed through the relationship keep partners together. Another is that a temporary better offer from outsiders may not be sufficient for someone who has already invested in relation-specific assets to leave the current relationship. Social and psychological assets, such as the warm memory of a pleasant past and mutual understanding, may be considered relation-specific assets that keep people in these relationships. Finally, commitment to a particular partner often reduces the level of trust in “outsiders” (see Kiyonari and Yamagishi, 1996, for experimental support), creating a vicious cycle of distrust of outsiders: those who do not trust “outsiders” tend to stay in committed relationships, and because they avoid “outsiders” they become even less trusting of “outsiders.” Yamagishi et al. (1998) further connect the tendency to form a committed relationship with the individual’s low level of general trust in others. They show in a cross-cultural setting (comparing the USA and Japan) that those who have high trust in others in general are less likely to form committed relationships. In Chapter 5, we follow up with a cross-cultural study (comparing the USA, China and the Netherlands), which shows that simple mere exposure is sufficient to increase commitment, even without an actual solution to the trust problem. Yamagishi et al. (1998) argue that general trust (or trust in people in general) provides a psychological springboard for people who have been “confined” to committed relationships to move out into the larger world of opportunities. However, as we argue in Chapter 6, general trust addresses only one of the concerns about switching to new partners. It mitigates concerns about social uncertainty, but not about resource uncertainty. On the psychological level, a different antidote is required for resource uncertainty, such as general optimism. We argue that general trust and optimism together serve as two mechanisms that help people to explore new relationships with strangers, thus decreasing commitment. Next to the uncertainty reduction mechanism, exchange theorists have recently started to recognize the importance of emotions in exchange commitments. Ed Lawler, the main proponent of the emotion argument postulates that in repetitive exchange, groups and relations become salient social objects that have a cognitive or subjective reality to actors (Lawler et al., 2000; Lawler,
10
Chapter 1. Introduction
2001). As such, these relations or groups may take on objective value and become ends in themselves (cf. Lawler and Yoon, 1996). Lawler and Yoon (1996) contend that success at exchange makes people feel good, while failure makes them feel bad. Their theory of relational cohesion states that individually felt emotions unleash a cognitive process through which the emotion is attributed in part to the relation or group that constitutes the context of the exchange. In this way, groups can become objects of intrinsic value to actors due to the positive emotions generated from exchange. Commitment in close relationships A special case of interpersonal commitment is close relationships, such as marriage and intimate partnership. Close relationships research has been the realm of psychology and social psychology, and so it is little wonder that it has identified the duality of emotional and rational explanations much earlier. Many studies in close relationships psychology refer to this duality as attraction and constraints (Adams and Jones, 1999). According to Goode (1960), the attraction (or “positive pull”) aspect is strong for example in romantic couples having a mutually satisfying and harmonious relationship. Both partners actively work together to ensure the future of the relationship. On the other hand, a constraining mechanism could similarly produce stay behavior. Even a marriage that exhibits no attraction anymore for either partner could nevertheless continue to exist due to external reasons, such as the sake of children’s well-being or to uphold appearances in a society where divorce is unacceptable (an “empty shell” marriage in Goode’s terms). Hinde (1979) creates a similar dichotomy when he distinguishes endogenous from exogenous commitment. Endogenously committed people strive to maximize the outcomes of their relational partner, even at the cost of their own interest. In contrast, exogenous commitment is based on the legal and social environment in which the relationship is embedded. In marriage commitment, Johnson (1973; 1991) introduces a third aspect by distinguishing between personal, structural (constraint) and moral-normative commitments. Personal commitment is the individual preference for staying in the marriage (because one wants to); structural commitment comes from avoiding negative consequences of the dissolution of the relationship (because one has to). Finally, moral-normative commitment arises from a sense of obligation, to do the right thing, to uphold personal behavioral consistency (because one ought to). A key psychological source of moral-normative commitment is the avoidance of cognitive dissonance – divorce may be in conflict with one’s view about marriage, or having made a public declaration through marriage vows. Another source is a sense of obligation to one another, regardless of what others think: one may want to remain true to the promise made in the wedding vow.
1.1. Background
11
From our perspective of an emotional-rational dichotomy in commitment, moral-normative commitment occupies a special position. It could be classified under rational explanations, simply as a factor that modifies instrumental properties of outcomes within a deliberative thought process. On the other hand, it could be part of an emotional explanation, inasmuch as norms are internalized and modify the emotional preferences of the individual. Within interpersonal relationship research, it is perhaps Rusbult who comes closest to establishing a rational choice framework for commitment. Building on Becker’s side-bet theory (1960) and Blau’s work on commitment (1967), Rusbult created an investment theory for interpersonal commitment (1980; 1983). According to the investment theory, the level of commitment to a relational partner is determined by multiple interconnected factors, such as relational satisfaction (the ratio of rewards and costs in the relationship), the quality and availability of alternatives or alternative states (e.g. singleness), and prior investment in the relationship. Having a highly rewarding relationship increases commitment, but so does not having satisfactory alternatives. Yet, in other works, Rusbult gives implicit indication that a rationality framework is insufficient to explain many aspects of interpersonal commitment. Johnson and Rusbult (1989) show, for example, that people unconsciously devalue potential alternatives the more committed they are to their current partner. Doing so, people distort key variables of a rational choice equation. Organizational commitment A large body of research studies commitment to organizations. In the discourse of organizational commitment, commitment refers to the attachment of a member or employee to an organization. It is sometimes used interchangeably with other concepts, such as cooperativeness and stay behavior, or even more broadly, organizational citizenship behavior (see Moorman and Blakely, 1995; Organ, 1988). Organizational commitment research is largely motivated by the insight that members who are more committed, will perform better and regard the interest of the organization as common with their own, are less stressed, and less likely to leave the organization. Meyer and Allen (1991) integrated many of the divergent conceptualizations and measurements of commitment into a coherent theoretical framework. Their model is based on the recognition that there are three main aspects (or “mindsets”) of organizational commitment: 1. Affective Commitment is the employee’s emotional attachment to the organization. It refers to identification with the goals of the organization and a desire to remain a part of the organization. The employees commit
12
Chapter 1. Introduction to the organization because they “want to”. In developing this concept, Meyer and Allen drew largely on Mowday et al.’s (1982) concept of commitment. 2. Continuance Commitment lies behind the commitment of an individual who perceives high costs of losing organizational membership (cf. the side bet theory, Becker, 1960), including economic losses (such as pension accruals) and social costs (friendship ties with co-workers) that would have to be given up. The employees commit to the organization because they “have to”. 3. Normative Commitment is created by feelings of obligation to the organization. For instance, the organization may have invested resources in training an employee who then feels an obligation to put forth an effort on the job and stay with the organization to repay the debt. It may also reflect an internalized norm, developed before the person joins the organization through family or other socialization processes, that one should be loyal to one’s organization. The employees stay with the organization because they “ought to”.
According to Meyer and Herscovitch (2001), an employee has a “commitment profile” at any point in time that reflects high or low levels of all three of these factors, and different profiles have different effects on workplace behavior such as job performance, absenteeism, and the chance to quit. These three factors are thought to jointly determine the overall level of an employee’s commitment to the organization Meyer and Allen (1991). Compare how similar this trichotomy is to Johnson’s model above (1973; 1991) under “Commitment in close relationships”.
1.2
Toward an evolutionary explanation
With the advent of sociobiology, and later the rapid growth of evolutionary psychology, many aspects of human behavior have been convincingly explained from an evolutionary perspective, relying on dynamics of genetic and cultural evolution. The major argument of evolutionary psychology (see Cosmides, 1989; Cosmides and Tooby, 1993) is that human ancestors spent a vast amount of time in a relatively stable environment of the Pleistocene, starting 1.8 million years ago and spanning until about 12,000 years ago. During the time spent in this ancestral environment, human brains and some of the most fundamental sociocultural institutions respectively, underwent a long adaptation process. During evolutionary adaptation (Darwin, 1859), the characteristics of an individual (trait) undergo random changes (mutation) that are inherited by
1.2. Toward an evolutionary explanation
13
their offspring. Through mutation new traits may appear, increase in strength or disappear2 . When the combination of traits (phenotype) of an individual increases reproductive success relative to other individuals, i.e. by increasing the chances of the individual surviving until a reproductive age, the traits of this individual become more prevalent in the population, through the relative increase in the number of offspring possessing the trait (natural selection). Traits that specifically increase mating opportunities, usually through some highly observable physical trait (e.g. the peacock’s colorful tail) may spread even faster (sexual selection). This process led to the stabilization of those cognitive abilities and social preferences which solved problems frequently encountered in our prehistoric ancestral environment. Due to rapid changes in our civilization in the last few millennia, many of these stable adaptations are no longer beneficial but nevertheless continue to influence the behavior of contemporary humans. One example is that, although an estimated 132,687 people sustain gunshot wounds that result in death or emergency treatment in the USA annually (Beaman et al., 2000), and only a handful of people are killed or injured by snakes and spiders, people learn to fear snakes and spiders roughly as easily as a pointed gun, and much more easily than an unpointed gun, rabbit or flowers (Öhman and Mineka, 2001). The explanation from evolutionary psychology is that snakes and spiders were a large threat in the ancestral environment but guns, rabbits and flowers were not. Several attempts have been made to construct a similar evolutionary explanation for commitment (in the general sense of promises and threats) that brings together the emotional and rational sides. As one of the main proponents of this line, Nesse (2001a), puts it: [There are] abundant examples of the importance of commitment in human social life. The evidence is so compelling that one cannot help but wonder why explanations for cooperation have been so narrowly dependent on methodological rationalism and individualism. I suspect the reason is the absence of a framework that can account for actions that seem irrational. In the framework of commitment, such behaviors are not only explicable, they are expected. Certain emotions seem opposed to reason because they are opposed to reason. In the short run they seem mysterious, but in the long run on average they give advantages that shape psychological traits that change the structure of human society. These psychological traits must be incorporated into our model, however difficult that may be (p. 161). 2 Note that according to the theory of cultural evolution (Boyd and Richerson, 1985) such an evolutionary process need not take place on a genetic level. They showed that culture can evolve by a very similar dynamic as genetically based traits evolve by natural selection. Culture also undergoes mutation, individuals have cultural offspring, etc.
14
Chapter 1. Introduction
The answer he proposes is to regard deep rooted emotions related to commitment (to an action) as evolutionary adaptations that serve a good purpose in general and in the long run but due to their hardwiring easily come in conflict with rational deliberation. He argues that the parts of the human brain that evolved latest in our history, the frontal lobes, closely match the abilities needed to use commitment strategies (p. 34). It appears that the frontal lobes are especially well-suited to calculating trade-offs between short-term costs of giving up options and long-term benefits that may or may not be obtained. Such calculations are inherently complex, because they involve considerations about social capital, and would be impossible without specialized mental hardware. According to Nesse, the frontal lobes are also involved in the ability to empathically identify with another person, which is essential to predicting whether the other will fulfill a commitment. The weakness of Nesse’s argument is that it attempts to cram too much under the explanatory umbrella of natural selection. In his book, he integrates works from psychology, game theory, ethology, law, medicine, religion and mythology. Doing so, his argument gets fragmented and lost in the myriad aspects of general commitment. In the end, some of the phenomena and mechanisms considered can only be linked to natural selection through smaller or larger jumps in the argument. In fact, Nesse tries to explain human cooperation and non-kin altruism arguing for a capacity for making threats and promises (commitment in the broad sense) in general, but his argument relies heavily on long-term relationships (commitment in the interpersonal sense). It could possibly strengthen his theory if the evolutionary argumentation were restricted only to the simpler and more specific idea of interpersonal commitment. Another proponent for the crucial role of emotions in commitment is Robert Frank. In his seminal book “Passions within Reason” (1988) he argues that social environments naturally produce situations where commitment could potentially play a pivotal role, yet there is little room for formal commitment devices, such as contracts or other tangible hostages. In these cases, the best solutions are emotional commitments. One of the social emotions Frank argues for, as a relatively hard-to-fake signal of commitment, is sympathy (Frank, 2001). Sympathy enables people to detect other’s emotional state and experience it to some extent. Detecting sympathy in others helps to make promises about future cooperation more credible. Another social emotion that makes commitments credible without tangible assurances is anger. In a world of purely rational self-interested people who have perfect self-control, all acts of defection where the costs of retaliation outweigh benefits would go unpunished. An angry person, however, seldom gets recognized as a rational one, leading to an increase in the credibility of his threat of punishment, and thus decreasing the expected benefits of defection
1.2. Toward an evolutionary explanation
15
in the first place (Frank, 1988).
1.2.1
Separating ultimate and proximate explanations
When attempting to construct an evolutionary explanation for any kind of behavior, it is important to separate parallel explanations on at least two different levels of causality, the proximate and the ultimate level (Mayr, 1961). Proximate explanations identify environmental stimuli that trigger mechanisms within the individual as the causes of physical expression of the behavior. For example, in answer to the question, “why do songbirds sing?”, one might argue that increased daylight in the spring leads to increased testosterone production which activates parts of the brain in male songbirds. This explanation identifies a proximate mechanism (a neurobiological one in this case) in response to a direct stimulus (increased sunshine) to explain behavior (singing). Such a proximate explanation, however, might leave one with a sense of unsatisfied curiosity. In order to answer why such a proximate mechanism came to exist in the first place, one needs to look for an ultimate explanation, on a more general level of evolutionary causation. The reason why male songbirds sing is that singing attracts females and defends territory from other males. Consequently, those males who sing have better chances of reproducing and spreading their habit of singing into the next generation of songbirds, than those males who do not sing. Such an ultimate theory has the advantage of explaining behavior, while at the same time encompassing and justifying the proximate explanation.3
1.2.2
How evolutionary theory helps to explain seemingly irrational behavior
There are at least three systematic4 ways in which evolved behavior may depart from the seemingly rational. The first two result from the fact that ultimate functions are implemented through proximate means, and the third is based on fundamental constraints on information processing. Proximate mechanisms are always imperfect in the sense that they were the first solution, discovered randomly by natural selection, which addressed a specific problem of survival and reproduction in the simplest and most costefficient way in a certain environment. As soon as there is a change in the environment, a proximate mechanism can easily lose its efficiency or even turn 3 Ultimate (also called holistic) explanations are subject to criticism by reductionists who claim that because ultimate explanations are functional, they lack a sufficient causal argument. We provide a counterargument to this criticism in Chapter 7, page 137. 4 By “systematic” we mean that behavior fails to be rational in the same way within the same context for a large number of individuals, i.e. not as isolated occurrences of some random or transient mistake in individual reasoning.
16
Chapter 1. Introduction
against the individual. Consider in the previous example the appearance of a human hunter who learns to imitate the calling of the male bird and thus easily captures female birds. In this case, females who have evolved a preference for males’ songs experience a serious decrease in their survival and reproductive chances. By definition, proximate mechanisms lead to stable behavior across different contexts. But while they create a clear adaptive advantage in one context, they could lead to maladaptive behavior in another. The first possibility for such errors is that a stimulus from the environment is falsely interpreted by the individual as a trigger for a proximate mechanism (a “false positive”, or “type I error” in statistics). The reason why the proximate mechanism could still be left in place by evolution is that the relative cost of the false alarm is smaller than the cost of not recognizing the real stimulus (a “type II error”). According to Error Management Theory (Haselton and Buss, 2000; Haselton and Nettle, 2006), humans acquired a large number of biases that increase the amount of false positives, when false negatives are extremely costly. An interesting example is that people develop a strong aversion to a certain kind of food, if its consumption was closely followed by sickness in the past (Garcia et al., 1966). This mechanism protected ancestral humans against consuming poisonous food sources (cf. Sripada and Stich, 2004). Such behavior could also be regarded as rational if information collection and processing are assumed to be costly. Therefore, the departure from standard rationality in this case is not so much the crude causal approximation between poisonous food and sickness but the fact that the aversion is manifest as a discomforting sensation in the gut, and not as the end-product of a deliberative thought process. An example with regard to interpersonal commitment is the laboratory studies carried out by Lawler and collaborators (Lawler and Yoon, 1993, 1996; Lawler et al., 2000). In these experiments, people became committed to their partners and reaffirmed their commitment with costly gifts when in fact they had never met these partners face to face and were ensured by the experimental setting that they never would. In this case, the bias for interpersonal commitment, a proximate mechanism, misfired in an inappropriate context (terminology from Sripada and Stich, 2004). The second way in which rationality could fail is when a signal is correctly recognized but the response given to it is no longer adaptive due to changes in the environment itself. A core assumption of evolutionary psychology (Cosmides, 1989; Barkow et al., 1992) is that the environment we live in today is radically different from the environment of evolutionary adaptation (ancestral environment). Therefore, some of the evolved stimulus-response mechanisms have become maladaptive. An example from the domain of interpersonal relationships is the very recent phenomenon of Internet addiction taking place among a worryingly large portion of ordinary people. The majority of these people turn to on-line chat
1.2. Toward an evolutionary explanation
17
rooms and role-playing games in search of social support, sexual fulfillment, and an opportunity to safely express forbidden aspects of their personalities. Adverse results include social withdrawal in the real world and loss of control, which are typical of other forms of addiction (Henry et al., 1997). In other words, people follow their evolved need for socialization but given the transformation of our social environment due to rapid technological development, the individual’s fitness is negatively affected. The third way in which human decision-making may depart from rationality is linked to information. In order to make rational decisions by choosing between different actions that lead to different outcomes, one needs information about these outcomes. If evolution ultimately favored rationality, it would also have favored mechanisms that help to obtain and process information accurately. There is mounting evidence that evolution sometimes works in the exact opposite direction. This is most notable in the case of evolved cognitive biases and optical illusions (see e.g. Haselton and Buss, 2000; Haselton and Nettle, 2006; Gigerenzer and Todd, 1999). Among the numerous examples, consider Evolved Navigation Theory. According to this theory, humans were selected to perceive physical characteristics of the environment (e.g. height and altitude) not as precisely as possible, but rather with a factoring in of the dangers they represent for individual fitness. Researchers in an experiment (Jackson and Cormack, 2006) asked one group of people to estimate the height of a very tall lookout point by looking at it from its bottom and another group to do the same from the top. It was found that people on the top consistently overestimated altitude, in proportion to the increased risk of falling. This shows how evolution can build safeguards into our cognitive apparatus that act against standard rational calculation. Consider now the finding of Johnson and Rusbult (1989) from earlier in this chapter, which shows that people systematically underestimate alternative partners, the more committed they are. If evolution ultimately favored the choice for a rational decision, it would have made sure that information about alternative partners is as accurate as possible at the time of making a decision. If, however, evolution aimed at stabilizing interpersonal commitments, it would have biased decisions in exactly this direction.
1.2.3
Constructing an ultimate explanation for interpersonal commitment
Although in his 1988 book Frank sets out to summarize empirical support for an evolutionary explanation for commitment in the general sense, many of his examples are more relevant for commitment in the interpersonal sense. Frank refers to marriage as a key example for a commitment dilemma. People search for the perfect mate, but settle for someone after a certain period of exploration despite knowing that there is certainly someone else out there, not
18
Chapter 1. Introduction
yet encountered, who would make a better spouse. And although a marriage contract may create a formal token of commitment, this is hardly the reason why people stop exploring further mates. A far more secure commitment is ensured by emotional bonds of affection (Frank, 2001). These emotional bonds ensure that even if someone kinder, better looking, or richer, who would originally have been preferred over the current partner, comes along now, the threat to the current commitment is diminished. But what does this have to do with evolution? There is growing acceptance among biologists of the idea that marital commitment is a key factor in enhancing the reproductive success of humans (Hrdy, 1999; Martin, 2003; Foley, 1996; Geary, 2000; Pillsworth and Haselton, 2005), indeed more so than in the case of any other primate species. In order to be able to pass through the birth canal of their mother with their large brain unharmed, human infants need to be born at an earlier developmental stage than other primate offspring (Hrdy, 1999). Consequently, they are more helpless and require substantially longer parenting (Martin, 2003). Therefore, finding a committed father who is present and cooperative during this extended period of parenting is instrumental for the reproductive fitness of humans5 (Foley, 1996; Geary, 2000; Pillsworth and Haselton, 2005). Indeed, there is a wealth of empirical findings in psychology and social psychology that gives further support for the existence of a consistently biased emotional-cognitive framework facilitating interpersonal relationships and commitment (cf. Baumeister and Leary, 1995). People in every society on earth belong to small primary groups that involve face-to-face, personal interactions (Mann, 1980). Festinger et al. (1950) found that mere proximity is enough for people to develop social bonds, and is especially suitable to compensate for differences in age or race (Nahemow and Lawton, 1975). Ostrom et al. (1993) showed that people memorize things related to close acquaintances on a person basis, whereas information related to looser contacts is stored and organized based on attribute characteristics (e.g. traits, preferences and duties). There is evidence that forgiving a misconduct of a committed partner directly enhances psychological well-being of the one who forgives (Karremans et al., 2003). It has also been shown that when people evaluate potential alternative partners, they unconsciously devalue potential alternatives the more committed they are to their current partner (Johnson and Rusbult, 1989). Kiyonari and Yamagishi (1996) give experimental support that those who stay committed to steady partners not only increasingly trust their partner, but 5 Indeed there is a possibility that next to natural selection, sexual selection also contributed to the proliferation of a commitment trait. Since human women need to find potentially committed mates to ensure the survival of their offspring, showing interpersonal commitment in social relationships in general could have served as a costly signal of males’ willingness and ability to become committed fathers.
1.2. Toward an evolutionary explanation
19
also increasingly distrust outsiders, leading to a “vicious cycle of distrust in outsiders”. On the one hand, a strategy of commitment appears to be efficient in forging beneficial relationships, yet it also loses out by letting potentially good alternatives slip away, and moreover, it gives way to exploitation within the relationship. To better understand these mechanisms and their interaction under a complex, evolutionary dynamic, we create formal (computational) models of the ancestral environment. Our models for the evolution of deeply rooted emotions underlying interpersonal commitment rely on a series of previous works by Henk de Vos and his collaborators (de Vos and Zeggelink, 1997; de Vos et al., 2001; Zeggelink et al., 2000). These researchers designed an agent-based computational model based on the following minimalistic assumptions about conditions of the ancestral environment: 1. People lived together in relatively small groups. 2. The environment was harsher, its impact less buffered, and resources more scarce than today. 3. In lack of many modern social institutions, help from fellow individuals was more important for survival than today. 4. The environment and subsistence technologies were more stable over an extended period of time than in modern civilizations, which made it possible for evolutionary pressures to hardwire preferences. De Vos and colleagues created a help exchange model, in which members of a relatively small group are dependent on the help of others to survive an event of distress from time to time. They compared two major contestants in their simulations of the evolution of exchange strategies, a strategy based on calculative reciprocal cooperation and a strategy based on commitment. De Vos and collaborators found that when each of the strategies competes against opportunistic players – i.e. actors who are unwilling to help but accept help from others – commitment is more viable than calculative reciprocity. De Vos et al. tentatively concluded from their computational experiments that under conditions of the human ancestral environment, selection pressures might have shaped a tendency towards commitment and largely unconditional cooperation. This tendency may still be present in contemporary humans, even though the pressures that formed it are weakened or no longer in place. However, their studies were strictly limited by the small number of strategy variations they examined. This presents a problem because overly cooperative agents following a commitment strategy could easily fall prey to smart cheaters, a possibility that their model could not account for. Moreover, as Binmore (1998) argued forcefully, the outcome of computer tournaments and simulations of evolutionary dynamics strongly depends on the set of strategies that are initially present in a population.
20
Chapter 1. Introduction
To address whether and to what extent these two potential problems reduce the viability of commitment, we propose in Part I (Chapter 2) a method to considerably and systematically enlarge the set of behaviors examined in the original analysis of the help exchange model. The core idea is to represent behaviors as determined by a set of individual preferences, or traits with respect to possible exchange outcomes. Agents in our model are boundedly and subjectively rational in the sense that they make decisions to cooperate, defect and change partners with the goal of maximizing subjective utility (or satisfaction) given their preferences. However, maximizing subjective utility based on individual preferences in our model does not necessarily lead agents to optimal exchange outcomes. We assume that individual preferences or strategies are subject to evolutionary pressure that selects for successful strategies based on the objective fitness consequences of the behavior resulting from the strategy. This approach is similar to the “indirect evolutionary approach” proposed by Güth and Kliemt (1998).
1.2.4
Proximate mechanisms for interpersonal commitment
Is there support for an ultimate explanation for commitment through a corresponding proximate mechanism? More precisely, do contemporary humans have a stable, hardwired tendency to become committed to their previous interaction partners in an emotional way when it is not in their instrumental self-interest? In Chapter 5 we empirically test the existence of such a potentially hardwired tendency for commitment through a series of cross-cultural laboratory experiments. When arguing for the evolutionary origins of any aspect of sociality, it is usually better to rely on cross-cultural data, in order to rule out cultural explanations. Since the environment of evolutionary adaptation mostly predates the break-up of modern cultures, adaptations associated with the ancestral environment should be present in all cultures. This is not to say, of course, that a cross-culturally stable phenomenon necessitates an evolutionary explanation, or that a lack of cross-cultural evidence rules one out. Culture intricately interplays with how people decide and behave, which itself has implications for biological evolution (Boyd and Richerson, 1985). The idea behind a proximate explanation for commitment is that through repeated positive interactions, people’s view of a committed relationship becomes systematically biased in comparison with a strictly instrumental perspective. When the relationship later takes a negative turn, this positively biased perspective for commitment makes stay behavior and cooperation more likely than otherwise expected. According to the mere exposure effect (originally described by Zajonc, 1968), when being repeatedly subject to a nonrepulsive stimulus, one develops a positive affect toward the stimulus. For example, the more we listen to the same piece of music, the more we appre-
1.3. Methodology
21
ciate it. We argue that such a mere exposure effect exists between long-term interaction partners. There is evidence, for example, that the more we see the same face, the more attractive we find it (Rhodes et al., 2001). What is even more interesting, is the finding that people also trust others more if they have been exposed to them more times, even in the absence of any interaction that could support actual inferences about trustworthiness (Moreland and Beach, 1992).
1.3
Methodology
Studies reported in this volume rely largely on two key methodological approaches, agent-based computational modeling and laboratory experiments. Below is a brief description of both, and an explanation of their usefulness for answering our research questions. Part I is aimed at testing an evolutionary theory that posits the stabilization of a commitment trait under selection pressures of the human ancestral environment. Its goal is to compare the strengths and weaknesses of commitment to other social preferences, such as calculative reciprocity (fairness). Drawing on earlier work in this domain (de Vos et al., 2001) and weighing the complexity of the modeling task, we decided to apply a type of social simulation to our problem, agent-based computational modeling (ABCM). Whereas social scientists usually model social processes as interactions among variables, ABCM studies interactions among adaptive agents who influence one another in response to the influence they receive (Macy and Willer, 2002). In ABCM, all modeling information about the properties of individual agents and their behavioral rules are transformed into a formal language (e.g. a computer program). Subsequently, the dynamics of the model, as well as conclusions on the macro-level can be deduced through step-by-step computation from given starting conditions (Flache and Macy, 2005). The advantages of ABCM are especially apparent when modeling dynamic phenomena in groups that are highly complex, non-linear, path-dependent, and selforganizing. The obvious advantage is that the explanation draws on local interactions among agents and not on predefined global characteristics of the group (Macy and Willer, 2002). A possible alternative methodology in Part I would be game theoretical (e.g. equilibrium) analysis. The benefit of such analysis is that it is able to provide more universal hypothesis tests that benefit from the strength of a mathematical proof. Its drawback is that given the complexity of our model, coupled with the evolutionary dynamic, this methodological approach seems unfeasible. This is also illustrated by the relative complexity of a brief analysis of a strongly simplified version of our model in Appendix A. In Part II, we aim to empirically test the evolutionary argument through
22
Chapter 1. Introduction
identifying peculiarities in decision-making in exchange relationships among contemporary humans. The majority of our hypotheses predict links between specific, well-defined conditions and exact, quantifiable measures of commitment. Therefore, strict control over conditions and measurement is indispensable. Accordingly, we decided to test our hypotheses in laboratory experiments with human subjects, using anonymous, computer-based settings. A possible alternative here would be to rely on representative crossnational surveys, or data collected among contemporary small-scale (e.g. hunter-gatherer) societies. Such secondary data analysis, however, would seriously restrict the scope of hypotheses that we could test.
1.4
Outline of chapters
The rest of this dissertation is organized as follows. Part I lays the theoretical foundation for studying interpersonal commitment using formal computation models. These three chapters investigate whether a commitment trait could have been adaptive under the conditions of the ancestral environment, with each chapter gradually putting commitment under a stricter test. More specifically, Chapter 2 examines whether improving on a known weakness of fair reciprocity (the main contestant of commitment in earlier work of de Vos et al., 2001) eliminates the competitive advantage of commitment. We create an agent-based model that is capable of incorporating previous models, and at the same time offers more flexibility and robustness to study the relative viability of commitment strategies. This chapter also tries to answer whether the exchange network structures, which are formed spontaneously in the simulated populations, help to explain the relative differences in viability. Chapter 3 puts the viability of commitment strategies under a stricter test by extending the previous ecological model through accounting for evolutionary dynamics of selection and mutation. This also means that the strategies examined here are no longer a priori invented and specified by the modeler, but emerge spontaneously through random walks in the strategy space. In comparison with previous work, this can possibly lead to the emergence of more sophisticated opponents that may take advantage of the weaknesses of commitment. A defining characteristic of interpersonal commitment is that one forgoes interaction with potentially better alternatives in favor of a long-term partner. Chapter 4 puts the emphasis on inequalities between potential interaction partners, and asks whether a preference for high-resource (or highly capable) others is more important than a preference for old partners (commitment). It also examines whether becoming committed to average or low-value partners undermines the efficiency of commitment. In order to do so we extend our previous model by introducing an inheritable trait for high-resource others,
1.4. Outline of chapters
23
and non-inheritable inequality in individual capabilities/resources. Part II turns to the experimental investigation of whether an ancestrally evolved commitment trait influences behavior in contemporary humans. Chapters 5 and 6 examine commitment in various exchange situations using laboratory experiments with human subjects. These two chapters take up the two major lines in explaining commitment in exchange: positive emotions and uncertainty reduction. In Chapter 5, we study whether people have a tendency to escalate commitment to previous interaction partners, when it is not in their self-interest. The purpose of this study is to test whether people have a cross-culturally stable emotional preference for previous partners that acts as a decision-making bias even in anonymous, economic setting. At the same time, in Chapter 5 we also find that uncertainty can decrease commitment, which leaves us with a new puzzle. To resolve the puzzle about the effect of uncertainty on commitment, in Chapter 6 we propose to refine the explanatory framework that has been used in the literature of commitment in exchange. More precisely, we identify an important assumption that previous works have left implicit, about the cooperative intentions of people, that qualifies the effect of uncertainty on commitment. In addition, Chapter 6 is also an empirical counterpart of the theoretical Chapter 4, which argued that a preference for previous partners (commitment) should be stronger among contemporary humans than a preference for highresource partners. Accordingly, we argue that partner selection situations can create not only social uncertainty (about trustworthiness) but also resource uncertainty, and that each type of uncertainty has its independent effect on commitment. At the same time, we contend that individual characteristics, such as general trust and optimism, may also influence the level of commitment. Finally, Chapter 7 contains a summary of results, with a general discussion of the findings, and an evaluation of the strengths and weaknesses of the work accomplished. We locate our work within past research and point to possible avenues of future research. Since Chapters 2, 3 and 4 have been accepted for publication in international peer-reviewed journals and a book, and the material that makes up Chapters 5 and 6 is currently under review, these chapters are kept in their original article format. As a result, some overlap may be detected between different chapters, especially in their empirical motivation and model descriptions. On the positive side, this leaves each chapter stand-alone and selfexplanatory.
Part I
An ultimate explanation
25
Chapter 2
The Competitive Advantage of Commitment1
Abstract
A prominent explanation of cooperation in repeated exchange is reciprocity (e.g. Axelrod, 1984). However, empirical studies indicate that exchange partners are often much less intent on keeping the books balanced than Axelrod suggested. In particular, there is evidence for commitment behavior, indicating that people tend to build long-term cooperative relationships characterized by largely unconditional cooperation, and are inclined to hold on to them even when this appears to contradict self-interest. Using an agent-based computational model, we examine whether in a competitive environment commitment can be a more successful strategy than reciprocity. We move beyond previous computational models by proposing a method that allows systematic exploration of an infinite space of possible exchange strategies. We use this method to carry out two sets of simulation experiments designed to assess the viability of commitment against a large set of potential competitors. In the first experiment, we find that although unconditional cooperation makes strategies vulnerable to exploitation, a strategy of commitment benefits more from being more unconditionally cooperative. The second experiment shows that tolerance improves the performance of reciprocity 1 This chapter is based on Back, I. and Flache, A. (2006). The Viability of Cooperation Based on Interpersonal Commitment. Journal of Artificial Societies and Social Simulation 9(1), http://jasss.soc.surrey.ac.uk/9/1/12.html.
27
28
Chapter 2. The Competitive Advantage of Commitment
strategies but does not make them more successful than commitment. To explicate the underlying mechanism, we also study the spontaneous formation of exchange network structures in the simulated populations. It turns out that commitment strategies benefit from efficient networking: they spontaneously create a structure of exchange relations that ensures efficient division of labor. The problem with stricter reciprocity strategies is that they tend to spread interaction requests randomly across the population, to keep relations in balance. During times of great scarcity of exchange partners this structure is inefficient because it generates overlapping personal networks so that often too many people try to interact with the same partner at the same time.
2.1
Introduction
The most prominent explanation of endogenous cooperation in durable relationships is reciprocity under a sufficiently long “shadow of the future” (Axelrod, 1984; Friedman, 1971). In this view, actors engage in costly cooperation because they expect future reciprocation of their investment or because they feel threatened by future sanctions for non-cooperation (Fehr and Schmidt, 1999; Fehr and Gächter, 2002; Falk et al., 2001). Roughly, these analyses show that even in a competitive environment with changing exchange partners, strategies that reciprocate cooperation with cooperation and defection with defection, such as the celebrated “Tit-for-Tat”, are far more successful than strategies that aim to exploit their opponents. Evolutionary game theory has demonstrated that if exchange relations persist long enough, cheaters are outperformed by reciprocators. This is because reciprocators benefit from ongoing mutually cooperative exchanges, while cheaters gain at best a short-term advantage at the outset of the exchange. This, however, cannot offset the longterm losses caused by the early disruption of the exchange relationship. It has been suggested that this reciprocity explanation of cooperation applies to a number of domains ranging from business ties between organizations to interpersonal relationships. However, recent empirical studies of cooperative behavior, in particular in interpersonal relationships, indicate that often reciprocity may be much less strict and actors much less intent on keeping the books balanced than the original reciprocity argument suggests. A short excerpt from Nesse (2001b) offers good examples: Perhaps the strongest evidence that friendships are based on commitment and not reciprocity is the revulsion people feel on dis-
2.1. Introduction
29
covering that an apparent friend is calculating the benefits of acting in one way or another. People intuitively recognize that such calculators are not friends at all, but exchangers of favors at best, and devious exploiters at worst. Abundant evidence confirms this observation. Mills has shown that when friends engage in prompt reciprocation, this does not strengthen but rather weakens the relationship (Mills and Clark, 1982). Similarly, favors between friends do not create obligations for reciprocation because friends are expected to help each other for emotional, not instrumental reasons (Mills and Clark, 1994). Other researchers have found that people comply more with a request from a friend than from a stranger, but doing a favor prior to the request increases cooperation more in a stranger than a friend (Boster et al., 1995). Moreover, there is solid empirical evidence indicating that people have a tendency to build long-term cooperative relationships based on largely unconditional cooperation, and are inclined to hold on to them even in situations where this does not appear to be in line with their narrow self-interest (see e.g. Wieselquist et al., 1999). Experiments with exchange situations (Kollock, 1994; Lawler and Yoon, 1993, 1996) point to ongoing exchanges with the same partner even if more valuable (or less costly) alternatives are available. This commitment also implies forgiveness and gift-giving without any explicit demand for reciprocation (Lawler, 2001; Lawler and Yoon, 1993). One example is that people help friends and acquaintances in trouble, apparently without calculating present costs and future benefits. Another, extreme example is the battered woman who stays with her husband (Rusbult and Martz, 1995; Rusbult et al., 1998). Since the seminal work of Axelrod (1984), a range of studies has used evolutionary game theory to refine the strategy of strict reciprocity and adapt it to empirical criticism. One line of work focused on the advantages of “relaxed accounting” in noisy environments (e.g. Kollock, 1993; Nowak and Sigmund, 1993; Wu and Axelrod, 1995). Broadly, these experiments confirmed the hypothesis that uncertainty favors “tolerant” or “relaxed” conditionally cooperative strategies that do not always retaliate after defection of an opponent. Kollock (1993), for example, found that in noisy environments (with mistakes and miscues), strict reciprocity is prone to needless recrimination that can be avoided by looser accounting systems. However, these studies cannot address the empirical phenomenon of commitment to long-term exchange partners, simply because they apply a repeated game framework in which there is no possibility to exit from an ongoing exchange in order to seek a new partner. A number of authors have explored variations of Tit-for-Tat that combine looser accounting under uncertainty with selective partner choice. Computational analyses of exit effects (Schüssler, 1989; Vanberg and Congleton, 1992; Schüssler and Sandten, 2000) put the role of the shadow of the future for
30
Chapter 2. The Competitive Advantage of Commitment
emergent cooperation into perspective. The route to emergent cooperation that these studies uncover is commitment of cooperators to cooperators, with the consequence of exclusion of defectors from relationships with cooperative partners. This is based on the principle “be cooperative but abandon anyone who defects.” When enough members of a population adopt this strategy, cooperative players stay in stable relationships, leaving defectors with no one but other defectors to interact with. As a consequence, defectors perform poorly and conditional cooperation thrives even under anonymity conditions where unfriendly players can hide in a “sea of anonymous others” (Axelrod, 1984, 100) after they “hit and run”. Considering more complex agent architectures, Schüssler and Sandten (2000) show that strategies that are to some degree exploiters may survive under evolutionary pressure but even then the most successful strategies will have the property of staying with a cooperative partner who turns out to be difficult to exploit. Other computational studies that include partner selection and arrive at similar conclusions are, for example, Yamagishi et al. (1994), or Hegselmann (1996) (cf. Flache and Hegselmann, 1999b). While previous work using evolutionary game theory could demonstrate the viability of relaxed accounting and commitment under certain conditions, it is doubtful whether this suffices to explain how humans may have acquired the deeply rooted emotions and behaviors related to interpersonal commitment that have been empirically observed. This is why de Vos and collaborators (de Vos et al., 2001; Zeggelink et al., 2000; de Vos and Zeggelink, 1997) extended theoretical models with assumptions from evolutionary psychology (Cosmides, 1989; Cosmides and Tooby, 1993). According to evolutionary psychologists, the way our mind functions today is the result of an extremely long evolutionary process during which our ancestors were subject to a relatively stable (social) environment. Individual preferences for various outcomes in typical social dilemmas stabilized in this ancestral environment and still influence the way we decide and behave in similar dilemma situations today. To model a stylized ancestral environment, de Vos and collaborators designed a help exchange game in which members of a relatively small group need the help of others to survive a situation of distress from time to time. More precisely, in their model agents come into distress at random points in time and then ask other members of the group for help. They compared two major contestants in their simulations of the evolution of exchange strategies, a strategy they called “keeping books balanced” (KBB) and a strategy called “commitment”. KBB corresponds to a strategy of strict reciprocity that is willing to help another actor but only as long as the favor is returned by the recipient as soon as possible. Otherwise, KBB will exit the relationship and seek new exchange partners. By contrast, commitment needs only a few successful initial help exchanges with a specific partner to become unconditionally co-
2.1. Introduction
31
operative to its partner further on. Broadly, de Vos and collaborators found that when both strategies need to compete against “cheaters” – i.e. actors who are unwilling to help but accept help from others – commitment is more viable than KBB under a large range of conditions. They conclude that in an environment where unpredictable hazards occur, KBB may be too quick to abandon exchange partners who get into trouble a second time before first reciprocating. As a consequence, a KBB player may often end up with no one willing to help it. A commitment player avoids this problem, because once committed to a cooperative partner it will not leave the partner in times of need and thus will benefit from future help from this partner when it experiences distress. De Vos et al. tentatively conclude from their computational experiments that under conditions of the human ancestral environment, selection pressures may have shaped a tendency towards commitment and largely unconditional cooperation that contemporary humans may still have, even when the pressures that formed it are no longer present. However, it is clearly an important limitation of these studies that only three possible strategies, KBB, commitment and cheating, have been taken into account and confronted with each other in a tournament approach. As Binmore (1998) has argued forcefully, the outcome of computer tournaments and simulations of evolutionary dynamics strongly depends on the set of strategies that are initially present in a population. The small set of strategies used by de Vos and collaborators may hide two potentially severe problems for the viability of the strategy of commitment. The first problem is the unconditionality of the strategy’s willingness to cooperate once it has been committed to a partner. This property obviously makes commitment highly vulnerable to exploitation by strategies that try to take advantage of its willingness to help. The second problem is that commitment may lose out in competition against more tolerant modifications of strict reciprocity. As the work by Kollock (1993) and others suggests, such modifications may avoid the major weakness of strong reciprocity to disrupt potentially cooperative exchanges too readily when problems occur. At the same time, such strategies also are less exploitable than commitment, because they eventually avoid being exploited by a partner who steadfastly fails to reciprocate help. To address whether and to what extent these two potential problems reduce the viability of commitment, we propose in this paper a method to considerably and systematically enlarge the strategy set used in the original analysis of the help exchange dilemma. The core idea is to represent strategies as a set of individual preference parameters, or traits with respect to possible exchange outcomes in a relationship. Agents in our model are boundedly and subjectively rational in the sense that they take decisions to cooperate, defect and change partners with the goal of maximizing utility from their preferences. However, maximizing subjective utility based on individual preference values in our model does not necessarily lead agents to optimal exchange
32
Chapter 2. The Competitive Advantage of Commitment
outcomes. We assume that individual preferences or strategies are subject to evolutionary pressure that selects for successful strategies based on the objective fitness consequences of the behavior resulting from the strategy. This approach is similar to the “indirect evolutionary approach” proposed by Güth and Kliemt (1998). Our approach allows systematic mapping of a range of individual variation in decision-making rules, e.g. variation in the extent of commitment or strictness of reciprocity. With this, we can carry out a stronger test of the viability of commitment than de Vos et al. (2001). We use our model to carry out two sets of simulation experiments designed to assess the viability of commitment in a larger set of potential competitors. For this, we take the original design of de Vos et al. as a starting point but systematically relax the assumption of unconditionality of cooperation in the first set of experiments. In the second set of experiments, we introduce and compare various degrees of relaxed accounting to reciprocity (“fairness”) strategies. In Section 2.2, we motivate and describe the model and our extensions. In Section 2.3, the computational experiments are reported. Section 2.4 contains conclusions and a discussion of our findings.
2.2
Model
Our model is based on a delayed exchange dilemma game, which is very similar to the one originally proposed by de Vos et al. (2001). The game is played by n agents in successive rounds. In the first round all agents are endowed with fi fitness points. In the beginning of each round, Nature selects a number of agents with a given individually independent probability Pd who experience distress and thus become in need of help from other agents in order to preserve their fitness level. These agents who are struck by Nature are the initiators of interactions. They ask others for help which is either provided or not. Providing help costs fh fitness points. Moreover, assuming that help giving is a time-consuming activity, each agent may only provide help once during one round; and only agents who are not distressed themselves may provide help. If a help request is turned down, the distressed agent may ask another agent for help but may not ask more than m agents altogether in the same round. If an agent does not manage to get help before the end of the round, it experiences fd loss in fitness. If the fitness level of an agent falls below a critical threshold fc , the agent “dies”, i.e. it is eliminated from the agent society.
2.2. Model
2.2.1
33
Modeling strategies
Agents in our delayed exchange dilemma face two different types of decision situations from time to time. If they are hit by distress, they have to select an interaction partner whom they believe most likely to be willing and able to help them. On the other hand, when they themselves are asked to provide help they have to decide whether to provide it and in case of multiple requests, whom to provide it to. Thus the mental model, or strategy, of an agent is represented as a combination of two sub-strategies: one for asking help and one for giving help. In previous studies by de Vos and others, behavioral strategies of agents were defined in natural language in terms of a collection of condition-action rules (e.g. for agent ai : if agent aj helped me before when I asked, then help him now) and then translated into a programming language. Even for simpler strategies several such decision rules had to be formulated, and this inherent arbitrariness limited the generalizability of the model. Our most important addition to these models is that we integrate them into a utility-based framework and provide in this way an efficient method to cover a large range of different strategies. In our model, when an agent has to make a decision, it calculates utilities based on some or all of the information available to it without the ability to objectively assess the consequences of the decision on its overall fitness2 . Moreover, we assume that actors are boundedly rational in the sense of being myopic, they evaluate the utility of an action only in terms of consequences in the very near future, i.e. the state of the world that obtains right after they have taken the action. This excludes the strategic anticipation of future behavior of other agents. Since different agents calculate utility differently, there is variation in behavior. Some of the behaviors lead to better fitness consequences than others. In turn, more successful agents have better chances of staying in the game and propagating their way of utility calculus to other agents, while unsuccessful ones disappear. Recent advances in psychological research into interpersonal relationships point to the influence of subjective well-being experienced when making certain relationship-specific decisions (Karremans et al., 2003). Unlike many applications of evolutionary game theory, we define utility calculus such that agents derive an emotional utility from features of a relationship, in addition to materialistic costs and benefits of help exchanges. This emotional utility can be interpreted as feelings and emotions, such as togetherness, belonging, sense of safety, identity, pride, etc., and the lack of it as loneliness, insecurity, shame, etc. We concentrate our modeling efforts on describing and analyzing this additional utility as a function of the history of help exchanges in a relationship. One of our main goals is to determine whether utility calculus based 2 In
certain cases these objective consequences may actually be impossible to foresee for the agents or even for the modeler.
34
Chapter 2. The Competitive Advantage of Commitment
on some form of commitment can lead to beneficial fitness consequences. In our delayed exchange game, agents have a very focused set of information available about their physical and social environment. They are aware of the fact that they got into distress, they follow the rules of the game (e.g. ask for help when in distress), and they remember previous encounters with other agents. This means that they know who and how often helped or refused them and who was helped or refused by them in previous rounds. The implicit assumption we make is that information about interactions between third-party agents is either not (reliably) available to the focal agent or is simply not taken into account in decision-making. We restrict the information available to agents from their earlier interactions to the following situation-specific decision parameters of an agent ai for each interaction partner aj (i 6= j): Definition 1 (Situation-specific decision parameters). EHij ERij AHij ARij
= number of times i helped j (ego helped), = number of times i refused j (ego refused), = number of times j helped i (alter helped), = number of times j refused i (alter refused)
As we mentioned above, agents face two different decisions situations. Accordingly, we define two independently calculated subjective utilities that agents use in these two decisions. The utility of donating that agent ai gains from helping agent aj is defined as a function of the situation-specific parameters: D D D D D + ehD Uij = Um i · EHij + eri · ERij + ahi · AHij + ari · ARij , D D D D where Um expresses materialistic costs of the interaction; ehD i , eri , ahi , ari are agent-specific parameters (or traits) for donation of agent ai that determine the weight of the situation-specific parameters in the total utility. In the actual implementation, every time an agent has to make a decision, there is also a probability Pe that the agent will make a completely random decision. This random error models noise in communication, misperception of the situation or simply miscalculation of the utility by the agent. Taking this random error into account increases the robustness of our results to noise in general3 . For simplicity, we define the utility as a linear combination of situationspecific parameters weighted by agent-specific parameters. The utility of seeking is defined in the same way, the only difference is that agents may put different weights on the situation-specific decision parameters than in the utility of donation: 3 See
more about the problems of involuntary defection in Fishman (2003) and agents getting stuck in mutual defection in noisy environments in Monterosso et al. (2002)
2.2. Model
35
S S Uij = Um + ehSi · EHij + eriS · ERij + ahSi · AHij + ariS · ARij ,
Before agents make a decision, be it help seeking or help giving, they calculate the corresponding one of these two utilities for each possible help donor or help seeker. In case of help giving, they choose a partner with the highest utility, if that utility is above an agent-specific threshold Uit . If the utility of all possible decisions falls below the threshold utility, no help is given to anyone. Otherwise, if there is more than one other agent with highest utility, the agent chooses randomly. As an addition to this rule, if an agent ai is asked to donate by another agent aj with whom ai has had no prior interaction (therefore all situationspecific parameters are 0), ai assumes that AHij = 1. In other words, agents behave as if a successful interaction has already taken place between them. Suspicious (non-nice) strategies can be defined by choosing the utility threshold (U t ) parameter so that without any prior interaction the utility of seeking or donation is lower than the threshold utility. In the case of help seeking, agents also choose a partner with the highest utility but there is no threshold, i.e. agents in distress will always ask someone for help. Using these rules, a strategy S in our model is described by the way utility is calculated. In other words, a strategy can be fully described by the two times four agent-specific parameters and the utility threshold. 4 By specifying ranges for agent-specific parameters we can easily define classes of strategies which correspond to basic personality types. For example, we classify a strategy S as belonging to the group of Commitment-type strategies, if the fact that previously help was received from a certain partner increases the utility an agent derives from donating help to or seeking help from that partner: Definition 2 (Commitment). ahD , ahS > 0 and er, ar = 0 This means that an agent ai of the Commitment-type derives more utility from choosing an agent aj as an interaction partner, the more times aj has helped ai in the past. This is true for choosing from both a group of help seekers and from possible help givers. This also means that the cooperativeness of a Commitment type is unaffected by the fact whether their help was previously refused by an interaction partner. If the utility threshold U t is zero or negative and ahD > U t , the strategy starts by cooperating, if able to.5 Otherwise, it behaves as “Suspicious Commitment”, i.e. it starts with defecting but after some cooperative moves of alter, it becomes cooperative. 4 Since we are interested in the perception of the strength of a relationship between agents rather D and U S are constant in all interactions. than the perception of objective costs, we assume that Um m 5 Note, that for the sake of simplicity in explaining the behavior of strategy classes, we will assume that the probability of a decision making error Pe = 0 throughout this section.
36
Chapter 2. The Competitive Advantage of Commitment
In the remainder of this section, we show how a range of further strategies can be defined with our method. For simplicity, we assume a utility threshold of zero, if not mentioned otherwise. The strategy type of Defection can be modeled with the assumption that it derives zero or negative utility from donation under all conditions: Definition 3 (Defection). ehD , erD , ahD , arD 5 0 and min(ehD , erD , ahD , arD ) < 0 If the utility threshold U t is positive and ahD < U t the strategy always starts by defecting. Otherwise, it only starts defecting after some initial rounds of cooperation. In general, we say that a strategy is a cooperator if at least one of its donation parameters is positive. In all other cases the strategy is a variant of the Defection type. Such a subset of Defection is AllD, which never helps others but when it is in need it randomly chooses others to ask for help: Definition 4 (AllD). Donation: ehD , erD , ahD , arD < 0 Seeking: ehS = erS = ahS = arS = 0 A much discussed strategy (type), especially in the experimental economics literature (see e.g. Fehr and Schmidt, 1999; Fehr et al., 2002) is Fairness.6 This is based on the observation that people may be willing to invest in cooperation initially but will require reciprocation of these investments before they are willing to cooperate further. On the other hand people following the fairness principle are also sensitive to becoming indebted, therefore they will be inclined to reciprocate if they are in debt. In other words, their most important aim is to have balanced relationships. Again, translating this strategy class into our framework is straightforward. Definition 5 (Fairness). Donation: ehD < 0, erD > 0, ahD > 0, arD < 0 Seeking: ehS > 0, erS < 0, ahS < 0, arS > 0 Agents belonging to the Fairness class deduce more negative utility from helping if they helped their partner in the past or if the partner refused them before, and will deduce more positive utility from helping if the partner helped them or if they refused to help the partner earlier. The twist here is that it is actually the ones that are most likely to be selected for giving help to that are different from those that are most likely to be selected for asking. Note moreover, that in case of a “Truly Fair” strategy, we would make the additional assumption about absolute values of traits such that |eh| = |ah| and |er| = |ar|. 6 Social
psychologists also refer to this type of behavior as equity (cf. Smaniotto, 2004).
2.2. Model
37
Suppose, for example, that an agent ai receives two help requests at the same time, one from aj , whom ai has helped twice before but from whom ai received help already three times. The other help request comes from a partner ak who helped ai three times and received help three times. A truly fair-minded person should in this situation help aj and not ak , and this is D exactly what follows from our implementation, because in this case Uij = D D D Uik − eh and eh < 0. Without making the assumption about absolute values, however, we are able to examine a larger class of “Fairness-type” strategies, such as “Tolerant Fairness” which increases credits (ah, er) more than it increases debts (ar, eh). Note that U t must be negative or zero for Objective Fairness, otherwise it requires more cooperation from its partner than it is willing to perform itself. Another way of relaxing the strictness of Objective Fairness is to decrease U t , which allows an asymmetry in favor of alter, in the amount of required reciprocation. For analyzing the individual rationality of cooperation we also define a trigger strategy, Grim Trigger. This strategy is the strictest form of cooperation, in that it permanently retaliates after its partner or itself defected and never cooperates again. Definition 6 (Grim Trigger). Donation: ehD = ahD = 0, erD < 0, arD < 0 Obviously, our approach allows generation of a much larger range of strategies than we discussed above. For our present analysis, it suffices to use these strategy templates but we will explore a larger variety of possible behavioral rules in future work.
2.2.2
Evolutionary dynamic
The heart of our model is an evolutionary dynamic that ensures the selection of objectively successful strategies (preferences). The dynamic we apply in our simulation is based on the replicator dynamics (Taylor and Jonker, 1978). Broadly, the replicator dynamics dictate that after a generation of genotypes (strategies) replicates itself, each different genotype will be represented in the next generation according to its relative success compared to other genotypes in the current generation. This way, unfeasible or self-harming preferences gradually become less widespread in the population, and give way to more “rational” preferences (see also Güth and Kliemt, 1998). To ensure that the size of the group remains constant throughout a simulation run, we apply the replicator dynamics in the following way. Whenever an agent dies, we create a new agent whose probability of belonging to a strategy S is equal to the proportion of collective fitness that is held by the group of agents belonging to S at the time of the new agent’s birth.
38
Chapter 2. The Competitive Advantage of Commitment
The evolutionary dynamic of our model is a strong simplification of the actual genetic reproduction that could have taken place in human evolutionary history. One argument for this simplification is to avoid the unnecessary overparameterization of our model. The central assumption we make is that better exchange outcomes of a strategy type translate into better chances for the propagation of that strategy. To capture this, there is no need to include individual level variables such as average and maximum number of children, age at giving birth etc., which are actually irrelevant for answering our research questions. Thus the great advantage of the replicator dynamics for our purposes is that it keeps the model of reproduction on the macro level. This also means that we only model the evolutionary selection of strategies but not mutation (see more under Discussion and Conclusions). With our explicit model of evolution we improve upon previous work of de Vos et al. (2001) in a number of ways. In their study they did not explicitly model a replication dynamic but instead linked independent tournaments to each other in order to map evolutionary trajectories. More precisely, the authors assumed that in a sequence of evolution the final average distribution of strategies at the end of one generation taken across a series of replications of that generation would also be the initial distribution in all replications in the next generation. This reduces repeatedly the distribution of individual populations to its average trajectory, which may entail a biased picture of the eventual distribution that arises. For example, unlike de Vos et al. (2001), we consider in our analysis also those simulation runs in which the entire population becomes extinct before a generation ends. These runs were originally disregarded by De Vos et al. This may have biased their results towards an overestimation of the survival chances of Commitment because only replications in which Commitment survived could have reached the end of a generation. Moreover, unlike previous work, our model does not suffer from the specification of a “cut-off” parameter, i.e. there is no fixed number of rounds after which we stop our simulations.7 In this way, we can be sure that the evolutionary dynamic reaches an equilibrium state where the population is homogeneous, and avoid biasing our results towards strategies that may be only initially successful.
2.3
Results
De Vos et al. (2001) examined two cooperative strategies, Commitment and Keeping Books Balanced (KBB), both playing against defectors. They showed that Commitment, which is largely unconditionally cooperative to those previous interaction partners who gave help at least once, had better evolution7 To
be precise, a simulation run ends when agents of a strategy have completely pushed out their opponents from population.
2.3. Results
39
ary success than the strictly reciprocal KBB under a large range of conditions. They tentatively interpret this result as evidence for the advantages of being unconditionally cooperative in an environment with scarce and uncertain opportunities to receive help. We argue that another conclusion may also be possible. It is plausible that being so unconditionally cooperative still makes Commitment more exploitable in comparison with conditional cooperators. The relative success of Commitment in comparison with KBB may rather be a result of KBB’s disadvantageous feature to disrupt relationships too readily when some mishap occurs. In order to test this possibility, we first conducted a simulation experiment to assess to what extent it makes a difference in Commitment’s success against defectors, when various degrees of unconditionality are compared. We did this by comparing four different types of Commitment each playing against Defection. We then turned to the possibility that a more tolerant, fairness-based strategy may be more successful against defectors than Commitment or strict fairness (KBB), in an uncertain environment. To test this possibility, we compared Commitment with more tolerant versions of Fairness, and we also compared fairness strategies that vary in their degree of tolerance to each other. Finally, while De Vos et al. measured and compared the individual success of cooperative strategies playing against defectors, they did not consider the possibility of an actual evolutionary invasion against Commitment by a conditional cooperator that is less vulnerable to exploitation by smart cheaters. We also provide results of this type below.
2.3.1
Simulation setup
Our goal with the simulation experiments was to compare different cooperative strategies with each other in terms of viability when there are initially some defectors in society. The most important indicator of a cooperative strategy’s viability was its success in resisting this invasion of defectors under evolutionary pressures of selection and reproduction. More precisely, within each simulation run, we started out with a group of multiple strategies. We allowed this mixed group to play the game for an extended period under an evolutionary dynamic, until only one strategy was present in the population. Within one experiment, we independently repeated such simulation runs from their initial state n times8 , until standard errors of measured variables became sufficiently small in order to be meaningfully interpreted. At first, we kept all environmental and model parameters constant and varied only strategy parameters from experiment to experiment. Even using our compact way of representing strategies (see section 2.2.1), we need to define strategies in a 9-dimensional space (using two times four weight parameters and a threshold parameter). Assuming that parameters and the 8n
= 2000 independent runs were usually sufficient.
40
Chapter 2. The Competitive Advantage of Commitment
threshold can only take 5 possible values (i.e. -2, -1, 0, 1, 2), we are left with a strategy space of 59 = 1953125 individual strategies. Fortunately, vast parts of this strategy-space yield similar behavior and thus can be classified under common concepts such as Defection, Fairness, Commitment, Trigger, etc. (see strategy types described above). For example, multiplying all traits and the utility threshold of a strategy S with a positive value will yield a strategy S 0 that behaves identically to S. More generally, as long as a transformation on the trait parameters does not shift the level of utility below or above the threshold for any given situation-specific parameter and does not modify the ordering of alters, it has no effect on behavior. Therefore, in the analysis that follows we will not vary absolute values of single traits, only traits in proportion to each other.9 To assess the robustness of results derived from the simulation experiments we conducted exhaustive sensitivity tests for all sensibly variable parameters. We report interesting deviations from typical results in section 2.3.4 below. For a list of all parameters see Appendix B. Initial parameters To determine interesting initial parameters for the simulation experiments and to reduce the parameter space that must be explored, we conducted a game theoretical analysis of a simplified version of the dilemma. Our goal was to identify the set of conditions that makes the choice for agents between purposeful defection and (conditional) cooperation as difficult as possible. If cooperation places an excessively high burden on agents, or conversely, if cooperation entails no real sacrifices, the model would hardly yield any interesting insights. To approximate the conditions under which cooperation is rational at all in the delayed exchange dilemma, we calculated expected payoffs in a simplified version of the game using trigger strategies. A trigger strategy behaves so that as soon as its interaction partner or itself defects, it falls back into a period of unconditional defection. The most severe version of trigger strategies is Grim Trigger, which never switches back to cooperation after its partner or itself defected. Even after its own unintended defection (i.e. due to being unable to help), Grim Trigger applies the most severe punishment possible in the game, permanent retaliation. If the sanction imposed by Grim Trigger cannot deter a rational player from unilateral defection, then no cooperative strategy can do so. As a consequence, there exists no Nash equilibrium – that is: a rational outcome – in which both players choose a conditionally cooperative strategy (see Abreu, 1988). The simplifications we make for the sake of the formal analysis 9 In other words, we are not covering exhaustively the entire parameter space, only a very large part of it. It may still be possible that there is a specific trait combination that is superior under some conditions to the ones examined.
2.3. Results
41
are that we reduce the group size to two and that we omit the evolutionary dynamics and the possibility of death due to low fitness. After solving this simplified dilemma situation (see Appendix A), we get a condition for the rationality of cooperation in the form of a relationship between the probability of distress, the cost for helping and the cost for not getting help: fh < fd (1 − Pd ) We used this result to adjust the most important initial parameters of the model. In other words, we always varied the probability of distress, the cost of help and the cost of not getting help in a way that the above inequality remained true. The actual parameters that we used to draw figures below are: Pd = 0.2, fh = 1, fd = 20, fi = 100, fc = 0, N = 25, Pe = 0.05, m = 2. For the entire set of parameter ranges that we tested in the experiments, please refer to Appendix B.
2.3.2
The unconditionality of Commitment
We started our experiments with comparing four different versions of Commitment playing against defectors. The necessary and sufficient condition for a strategy S to be classified under commitment is that its ahD and ahS traits are positive, which means that agents belonging to S will be inclined to choose those alters for cooperation who have helped them in the past. In its simplest form, this is all that Commitment cares about, expressed by e.g. the following traits: [0, 0, 1, 0|1].10 We will refer to this strategy as Weak Commitment from now on. An important question about the behavior of Commitment is whether the fact that ego helped alter before (EH) should also increase ego’s willingness to cooperate. Intuitively, such a preference makes an agent more vulnerable to exploitation and holds no obvious benefits for ego, if it has an effect at all. Therefore, the second version of Commitment we examine has a non-zero value on its eh parameter: [1, 0, 1, 0|1], we denote it Strong Commitment. If we compare now how Weak Commitment and Strong Commitment play against Defection, we see that indeed the eh trait makes a difference. Whereas Weak Commitment managed to eliminate Defectors from the society in 5.1% of all simulation runs, Strong Commitment did so in 67.7%. These results are based on 2000 independent replications of the simulation for both conditions. What is important here are not the actual percentages but the relative success of Strong Commitment compared to Weak Commitment when playing against Defectors. While the survival statistic will shift in favor of defectors when 10 A strategy is described by four trait parameters for donation, four trait parameters for seeking and the utility threshold: [ehD , er D , ahD , arD |ehS , er S , ahS , arS |Ut ]. If parameters for donation and seeking are equal, we provide only four trait parameters instead of eight. Find more information above, in section 2.2.1.
42
Chapter 2. The Competitive Advantage of Commitment
the cost of giving help (fh ) is increased or the cost of not getting help (fd ) is decreased, Strong Commitment remains superior to Weak Commitment. Figure 2.1 shows how the fitness proportion of Weak and Strong Commitment agents playing against Defectors changes over time. In both cases, Defection initially owns 20% of the total societal fitness (and group size), Commitment 80%. Each curve corresponds to an independent simulation run. All runs end with one strategy completely outnumbering the other but each run may be of different length. Shorter runs are complemented with dashed lines for clarity. What is important to observe is the proportion of curves ending in 0% compared to those ending in 100%11 . Additionally, a black curve shows the average fitness proportion held by a strategy at each point in time across multiple runs. Note also how simulation runs tend to last much longer in the case of Strong Commitment. This indicates that it takes less time for Defection to push out Weak Commitment, than it takes Strong Commitment to push out Defection. Although Defection starts from a smaller proportion than its opponent in both cases, it clearly outpowers Weak Commitment in most runs. In the second case, by contrast, Defection hardly ever manages to climb to the fitness level of Strong Commitment. To assess the relative importance of the eh and the ah traits, we examined two “mixed” versions of Commitment: [1, 0, 2, 0|1] and [2, 0, 1, 0|1]. The former version represents a strategy that derives more utility from receiving help than from giving help to a particular partner, whereas the latter version derives more utility from giving than from receiving. Both had high survival statistics. The proportion of replications in which the corresponding Commitment strategy became universal in the simulated group was 64.2% and 72.4%, respectively. The results also hint at a stable positive effect of eh on survival success. Unconditionality and AllC One doubt that might have surfaced in the heedful reader about the characteristics of our model is that the more cooperative a strategy is, the more successful it will become due to the relatively low costs of cooperation. In order to show that this is not true, we provide the results for the strategy AllC playing against Defection. AllC is the upper end of cooperativeness: it always chooses to cooperate. All other strategies are either equally or less cooperative than AllC. If more cooperativeness implied higher survival chances, AllC should be the winner of all. 11 Note that in order the maintain the clarity of the graphs, we reduced the number of simulation
runs actually plotted on the figures.
2.3. Results
43
Figure 2.1: Weak and Strong Commitment playing against Defection This is not what we see in the results: Defectors managed to overthrow AllC in 88.6% of all simulation runs (see also Figure 2.2). Comparing the survival rates of AllC with those of any version of Commitment except Weak Commitment, it is obvious that AllC is less successful. The weakness of AllC is caused partly by its partner selection behavior. Since AllC is blindly cooperating it is also blind in choosing its interaction partners. When deciding who to give help to and who to ask help from, AllC is indifferent between all possible partners.12 The main weakness of AllC compared to Commitment is its lack of an explicit partner selection strategy. Both versions of Commitment are more likely to help those others who have helped them before. Accordingly, a player who tries to exploit Commitment will always be less likely to get help than 12 Therefore, it is important to define AllC as [0, 0, 0, 0|0] and not e.g. as [1, 1, 1, 1|1] or [2, 2, 2, 2|2] since in the latter two cases it would become a variant of commitment (i.e. it will attribute more utility to those he interacted with more times in the past). While the latter two definitions are perfectly identical to each other, the first definition prescribes surprisingly different behavior with regard to partner selection.
44
Chapter 2. The Competitive Advantage of Commitment
Figure 2.2: AllC playing against Defection
somebody else who cooperated with AllC, all other conditions being equal. Due to its random partner selection method, AllC is the upper end of not only cooperativeness but of unconditionality as well.
Commitment in comparison with fair conditional cooperators The classic fair type (known e.g. from Fehr and Schmidt, 1999; Fehr et al., 2002) keeps a close watch on the balance with its opponent. A fair ego calculates the balance with regard to help donation to a particular alter as follows. Whenever ego helps alter or alter refuses ego, ego subtracts a unit from the balance with alter and whenever alter helps ego or ego refuses alter, ego adds one unit to the balance. The balance with regard to help seeking is calculated in exactly the opposite way. That is, the balance is most favorable with respect to the agent who owes ego the most. In terms of our model, this strategy is defined as [−1, 1, 1, −1|1, −1, −1, 1| − 1]. We will refer to this as Objective Fairness. If we examine how this objective version of Fairness plays against Defection, we see that it has a 2.5% chance of surviving, which is lower than the worst we saw for Commitment. The failure of Objective Fairness (Figure 2.3) can be explained with the large number of rejections out of unwillingness to help. These rejections are due to asymmetries in the number of times helping partners become distressed: e.g. if ego becomes distressed too often compared to alter (i.e. before he can reciprocate help from alter), alter will no longer provide help for ego. This result is consistent with previous game theoretical and simulation analyses that pointed to the disadvantages of strict reciprocity in uncertain environments (e.g. Kollock, 1993). Objective Fairness can be straightforwardly modified in such a way that
2.3. Results
45
Figure 2.3: Objective Fairness playing against Defection
it becomes more tolerant against temporary fluctuations in the frequency of needing help. For our simulation, we define a corresponding strategy of Tolerant Fairness as [−1, 2, 2, −1|2, −1, −1, 2| − 1]. The survival statistics of Tolerant Fairness (see also Figure 2.4) are clearly better than those of Objective Fairness. Tolerant Fairness stayed standing against Defection in 13.2% of all simulation runs. This result, however, is still worse than the result of Strong Commitment.
Figure 2.4: Tolerant Fairness playing against Defection
46
2.3.3
Chapter 2. The Competitive Advantage of Commitment
Explanation: the importance of strong ties
To understand the variation in the success of different cooperative strategies, we first tested how viable they are when they play without defectors, in a homogeneous society composed of only one cooperative strategy. Using identical parameter settings and group size we found that there is still significant variation in success, even between unconditionally cooperative strategies. When Commitment-type agents play against each other, being largely unconditionally cooperative, the most frequently observed type of refusal is when an agent is not able to help another agent. Hence, whether strategies cooperate or not cannot explain variation in success of different commitment types. However, whom players select to cooperate with or to request cooperation from turns out to be decisively different between versions of Commitment. But why should partner selection matter when everybody else is playing Commitment and will never refuse to help out of unwillingness? There are two problems faced by an agent even in this homogeneously friendly world of fellow cooperators. One is if help is sought from an agent who is distressed himself and cannot help; the other is if multiple agents ask the same agent to help. A collectively ideal strategy works so that help requests are evenly distributed among agents who are not distressed. It turns out that commitment comes very close to being such an ideal strategy, due to a phenomenon of dyadization. To understand dyadization, let us consider what happens in the first few rounds. According to individually independent random events with probability Pd , a fraction of the whole society N · Pd becomes distressed. Since nobody has had an interaction before, the distressed agents will all choose randomly whom they ask for help. For a distressed agent the probability of losing fitness by the end of the run is composed of two parts: Pf itnessloss = P1 + P2 , where P1 is the probability that another distressed agent also asked and got help from alter, and P2 is the probability that alter is also distressed. Let us assume now that eh = 0 for all agents. If an agent ai who was not distressed and helped in the previous round becomes distressed now, that agent ai will face a population of equally preferred others to ask for help. Therefore, ai will have to choose randomly, facing the same probability P1 that the other unlucky ones faced in the previous round. Now, if eh > 0, all agents who gave help before will have a preference for those they helped, and thus P1 will be reduced in the second round. To put it more generally, in later rounds, if everybody simply chooses the partner they had the most interactions with, it is likely that there will be few collisions between help requests. This leads to the formation of increasingly
47
2.3. Results
strong ties, and this is what we referred to as dyadization. Graphing the social network of Weak and Strong Commitment clearly shows the difference in the extent of dyadization (see Figure 2.5; the darker a tie the more interactions have taken place between the two agents). In the case of Strong Commitment, we see few but stronger ties (“close friendships”), while in the case of Weak Commitment we see many but weaker ties. A more careful look reveals that most Strong Commitment agents have exactly one strongest tie (“best friend”), while this is less true for Weak Commitment and even less for Fairness (Figure 2.6). What we see in a network of Fairness players are homogeneously weak relationships and nodes with a larger degree in terms of the number of ties. All networks are graphed after 200 rounds, which means that the added strength of all ties (network density) in the networks is roughly13 equal. To precisely measure the dyadization of a network, we first compute the “individual dyadization” for each node in the network by dividing the strongest tie of a node by the total strength of all ties of that node. The strength of a tie is obtained by counting the number of interactions that have taken place along that tie. We average the resulting values over all agents to obtain the dyadization measure of the network. A network dyadization of 1, or perfect dyadization, means that every node is connected to exactly one other node. The dyadization after 200 rounds averaged across 2000 runs for a network of Weak Commitment players is 0.29, whereas it is 0.39 for Strong Commitment. The dyadization measure for Tolerant Fairness is 0.23. Invasion of cooperators on cooperators In a final test of the relative viability of Commitment compared to Fairness we let Strong Commitment play against Tolerant Fairness in the presence of Defectors. Commitment and Fairness both started with equal initial proportions (40% each) while Defectors were in a minority (20%). Although Commitment did better (46.3%) than Fairness (2.7%), it was Defection who won (51.0%) this tournament. Apparently, the cooperative group in this case was considerably weakened by the occasional lack of cooperation of Fairness players. Observing the network structure of all agents we found again the characteristic strong friendships between Commitment players. Defectors, at the other extreme, attempted to interact with as many others as possible, in search of partners to be exploited. Moreover, we also found some strong ties between Commitment and Fairness agents. What happens in these relationships is that the Commitment player becomes attached to the Fairness player after some initial rounds of helping. The problem for the Commitment agent arises if the relationship becomes unbalanced – Commitment will keep asking its Fairness “friend” for help even in the face of repeated refusals. This points to a weakness of Commitment. 13 Depending
on the actual number of stochastically arisen distresses.
48
Chapter 2. The Competitive Advantage of Commitment
Figure 2.5: Emergent social network of Weak and Strong Commitment players after 200 rounds
2.3.4
Sensitivity to initial parameters
To assess the robustness of the results reported above, we examined their sensitivity to the choice of initial parameters. The cost of helping, the cost of not getting help and the probability of distress As we pointed out earlier, the ratios of the three parameters, cost of helping (fh ), cost of not getting help (fd ) and probability of distress (Pd ) are essential to determine the individual rationality of cooperation. Increasing fh in proportion to fd results in better survival chances of defectors. According to the analysis of the simplified dilemma (see Section 2.3.1), increasing the probability of distress makes the conditions under which conditional cooperation is individually rational more restrictive. This is not what we see in the simulations. Increasing Pd from 0.05 to 0.2 actually benefits cooperators, especially Commitment. This result replicates what was found by de Vos et al.
2.3. Results
49
Figure 2.6: Emergent social network between Tolerant Fairness players after 200 rounds
(2001). The explanation is that Commitment players find each other sooner and strengthen their relationships faster, the harsher the environment is, i.e. the larger the probability of distress is. Initial distribution and group size In those simulation experiments reported above, where a group of cooperators played against a group of defectors, we started from an initial population of 20 cooperators against 5 defectors. Decreasing the initial population size to 10 or increasing it to 50, keeping the initial ratio of cooperators to defectors constant, did not result in notable deviations from our results. Changing the initial proportions of cooperators and defectors did change percentages of survival to some extent but it did not reverse our qualitative conclusions. More precisely, although we found that in all experiments increasing the initial proportion of defectors led to lower survival chances for cooperative strategies, the survival chances for strong commitment were still higher than those of weak commitment or fairness strategies. Moreover, decreasing the initial proportion of defectors did not change results qualitatively either because we found that the evolutionary dynamic enabled even a smaller but superior invading strategy to take over the entire population. We can conclude that the choice of these parameters does not affect our results qualitatively. Probability of decision-making error We found no unexpected results when varying the error of decision making between 0.0 to 0.5: the larger the error was, the smaller the difference be-
50
Chapter 2. The Competitive Advantage of Commitment
came between the behavior of different strategies as they all approached a completely random strategy. Defectors suffered most from a high level of decision-making error: although they became more marginally cooperative, their choice of when and with whom to interact was completely haphazard. Number of subrounds and initial fitness By increasing the number of subrounds (m), agents in distress have a higher chance of finding a helping partner within one round. The general effect we expected was that survival becomes easier for all agents, and that connections build up faster as agents encounter more alters during the same number of rounds. It is intuitively not obvious, however, who benefits more from a second (third etc.) chance – cooperators or defectors? Rerunning the simulations, keeping all parameters unchanged except varying m between 1,2 and 3, we found that it is defectors who have better chances of pushing out cooperators, the larger m is. Whereas more subrounds give more chances to get help within one distress period, higher initial fitness keeps the agent alive across multiple distress periods. We expected the same general effect for changing the initial fitness parameter (fi ) as for changing the subrounds parameter because an increase in either parameter results in increased survival chances of all agents. Varying the initial fitness parameter fi , we found an interesting nonlinearity. Increasing the initial fitness from 50 to 100 resulted in defectors becoming more successful against both conditional and unconditional cooperators. Further increasing fi , however, resulted in defectors becoming relatively even less successful against Commitment players than at lower initial fitness. The general result of Commitment being more successfully than Fairness when playing against defectors remained constant throughout this test as well.
2.4
Discussion and Conclusion
Our results suggest that strategies following some form of commitment behavior are highly successful under a wide range of conditions. Broadly, commitment is modeled as the extent to which cooperativeness with a particular partner becomes unconditional after some initial cooperative actions of the partner. Counterintuitively, the faster an agent is inclined to solidify its relationships (see Strong Commitment), the less prone it is to exploitation. The reason is that a relationship between two Strong Commitment agents is built up – probabilistically – at least twice as fast as a relationship between a Strong Commitment and a Defector agent.
2.4. Discussion and Conclusion
51
Our approach shows that the success of Commitment remains stable even when a much larger range of strategy variation is allowed than in the previous computational experiments of de Vos et al. (2001). We find that strategies that base their behavior on fairness principles generally perform much worse than commitment strategies. A truly fair strategy suffers from its lack of tolerance when interacting with its own kind in an unpredictably “unfair” environment, where imbalances in the exchange cannot be avoided due to uncertain hazards. A more tolerant strategy that is nonetheless based on preferences for fair outcomes proves to be more viable. An interesting result of our model is that strategies that take their own past behavior into account – not just that of their interaction partners – make more successful decisions in general. To explicate the reasons for the success of commitment, we also studied the spontaneous formation of exchange network structures in the simulated populations. It turned out that commitment strategies derive a large part of their success from efficient networking: they avoid overloading few designated individuals with interaction requests and instead spontaneously create a structure that ensures an efficient coordination of help requests and help provisions. The problem with fair strategies is that they are predisposed to keeping their relationships in balance so that agents tend to spread interaction requests randomly across the population. During times of great need, this structure is inefficient because fair strategies in small groups generate overlapping personal networks so that often too many people try to interact with the same agent at the same time. While we tentatively conclude that our results support the hypothesis that commitment strategies are evolutionarily viable, we are also aware of a range of potential limitations of our analysis, some of which point in the directions for future research. A first objection to our study might be that we excluded the influences of reputation mechanisms on the relative success of strategies. There are a number of game theoretical studies and agent-based simulations that show how reputation mechanisms can sustain cooperation, because they help cooperators to effectively identify and punish cheaters (see e.g. Takahashi, 2000; Raub and Weesie, 1990). In our analysis, reputation effects were explicitly excluded with the assumption that agents rely only on their own experience about others when making decisions. We argue that taking reputation into account may be an unnecessary complication that would not lead to a qualitative change in the outcome of our comparison of Commitment with other versions of conditional cooperation. There is no particular reason to believe that Commitment would benefit less from reputation than other cooperative strategies. On the contrary, as Commitment players build up ties more readily and never abandon them afterwards, it may be particularly useful for Commitment players to use thirdparty information to identify in advance who is a reliable helping partner and
52
Chapter 2. The Competitive Advantage of Commitment
who is not. A further possible limitation of the present analysis is that we have not yet explored more sophisticated cheating strategies. It is possible that more sophisticated cheating may indeed undermine the viability of Commitment. The Defection strategy derived from AllD that we used in our experiments is not capable of taking advantage of what may be the most decisive weakness of Commitment, its inability to strike back once it is exploited by a partner to whom it has become committed. The only way Commitment punishes opportunists seeking help occasionally is by giving higher priority to friends with whom more numerous successful interactions have taken place. Similarly, Commitment will not even try to get help from its occasional interaction partners, if it has long-standing partners. In other words, instead of detecting cheaters by the number of times they defected, Commitment detects cooperators using the number of successful interactions. Nevertheless, there may exist more viable strategies outside the range that our analysis has covered. In future research we plan to extend our analysis in this direction. In particular, the effect of more sophisticated cheating strategies can be tested by allowing mutations to randomly generate strategies from a large set of possibilities. Clearly, this requires the modification of the evolutionary dynamic and significantly larger computational power than we used for the present study. A final line of future work may follow from resolving a simplifying assumption we made, to ignore possible differences between group members in terms of their attractiveness as exchange partners beyond their strategy. Such differences may for example come from variation in physical strength or more or less favorable local living conditions. Studies by Hegselmann (1996) (cf. Flache and Hegselmann, 1999b) suggested that variation in attractiveness may give rise to core periphery network structures in which the strongest population members exchange help with each other, driving weaker actors to the margin of help exchange networks. However, this work relied on conditionally cooperative strategies that resemble the strategy of Fairness. Accordingly, it is unclear whether variation in individual attractiveness may affect the viability of commitment strategies and also whether commitment strategies would give rise to the exclusion of weak members from exchange networks in the same way as has been found for Fairness-like strategies.
Chapter 3
The Evolutionary Advantage of Commitment1
Abstract
Why are people inclined to build friendships and maintain durable, nonreproductive relationships? Previous computational modeling work showed that it can be an efficient survival strategy to choose interaction partners based on relationship length, even if, as a consequence, individuals become unconditionally cooperative in long-term relationships (interpersonal commitment). Such committed individuals can outperform conditional cooperators who play in a fair, reciprocal manner (e.g. Tit for Tat). However, previous studies did not conduct a sufficiently strict test of the viability of commitment because they did not account for exploiters who specifically take advantage of the tolerance of commitment players. We allow this by extending previous studies with the possibility of randomly mutating strategies under evolutionary pressures, and thus give a much larger coverage of an infinite strategy space. Our results point to the lack of stable strategies: we find that emerging populations alternate between temporarily stable equilibria. We also show that the viability of strategies increases with increasing levels of interpersonal commitment, and that the effect of interpersonal commitment on viability is larger than the effect of fairness. 1 This
chapter is based on Back, I. and Flache, A. (Forthcoming). The Adaptive Rationality of Interpersonal Commitment. Rationality and Society.
53
54
3.1
Chapter 3. The Evolutionary Advantage of Commitment
Introduction
Among the species of the earth, humans exhibit the highest level of cooperation between genetically unrelated individuals (Gintis, 2003). Arguably, cooperation is the de facto key to our evolutionary success. At the same time, cooperation is problematic to explain from a rational actor perspective. Selfinterested actors often face a “social dilemma” (Dawes, 1980) where the rational pursuit of individual interests may lead them towards defection, while this in turn entails collectively undesirable outcomes. Game theory has identified repeated interaction as an important solution for the problem. In a world of harsh competition, repeated encounters reduce uncertainty about the trustworthiness of interaction partners (shadow of the past) while at the same time they create a strategic incentive for cooperation (shadow of the future) (Friedman, 1971; Axelrod, 1984; Buskens and Raub, 2002). Thus durable relationships are expected to be a hotbed of cooperation even in the absence of central enforcement because they provide incentives both to trust others and to honor others’ trust. From this perspective, it is hardly surprising that rational incentives to become committed to long-term cooperative exchange partners are particularly strong in uncertain environments (Schüssler, 1989; Kollock, 1994). A reduction in uncertainty is often more valuable than a probabilistic increase in payoff from a potential new partner, especially if switching itself is risky, costly or alternatives are scarce. This can explain why in situations where uncertainty may otherwise preclude the desirable outcome of mutual cooperation, social actors often restrict their own freedom of action by using commitment devices such as posting a hostage (Raub, 2004). However, this rational explanation of interpersonal commitment behavior is hard to reconcile with the empirical evidence that people tend to stay committed to long-term interaction partners even when (1) alternatives are available, (2) switching costs are low, and (3) uncertainty is of less concern. A growing body of empirical findings from both interpersonal relationships research (Karremans et al., 2003; Wieselquist et al., 1999) and exchange experiments (Kollock, 1994; Lawler and Yoon, 1993, 1996) shows that people have a tendency to remain cooperative with interaction partners who are occasionally uncooperative. Moreover, people tend to keep exchanging with the same partner even if more valuable (or less costly) alternatives are available. Such commitments also imply forgiveness and gift-giving without any explicit demand for reciprocation (Lawler, 2001; Lawler and Yoon, 1993). People help friends and acquaintances in trouble, apparently without calculating present costs and future benefits. Another, extreme example is the case of battered women who stay with their abusive husbands (Rusbult and Martz, 1995; Rusbult et al., 1998). In this paper we seek an explanation by following the general lead of the
3.1. Introduction
55
“indirect evolutionary approach” (Güth and Kliemt, 1998), which posits that individuals act rationally in the light of their preferences but also assumes that in the course of biological and cultural evolution individuals with social preferences and emotions – e.g. for fair distributions (cf. Bolton and Ockenfels, 2000), or for altruistic punishment (cf. Fehr and Gächter, 2002) – may have had a selective advantage, because their preferences produce more viable outcomes than those of pure egoists. This approach aims to integrate the endogenous explanation of such non-selfish preferences with the classical rational choice assumption that humans act (boundedly) rationally, given their preferences. The core idea is that preferences are embodied in the genotype and guide individual actions. Subjective preferences may be harmful to the individual or to the population but genotypes are selected on the basis of objective consequences of the actions that preferences produce. As preferences undergo selection and mutation, unfeasible and harmful preferences gradually become less widespread in the population, giving way to more “rational” sets of preferences. However, while the indirect evolutionary approach has proven to be a fruitful avenue in explaining phenomena such as emotional commitment to a certain course of action, altruistic punishment, or cooperation in the production of collective goods (Frank, 1988; Güth and Kliemt, 1998; Güth and Ockenfels, 2002; Gintis, 2003), the phenomenon of interpersonal commitment has received relatively less attention. Recently, some authors have begun to use the indirect evolutionary approach to explain interpersonal commitment (Back and Flache, 2006; de Vos et al., 2001; Zeggelink et al., 2000; de Vos and Zeggelink, 1997). While these analyses suggested that commitment might have been evolutionarily viable, we argue that the tests they used were not strict enough. De Vos and his collaborators argued in a series of papers that in a stylized “ancestral environment” a strategy based on commitment behavior could outperform a strategy based on calculative reciprocity when both strategies are in competition with one defecting strategy. The researchers modeled commitment as unconditional cooperativeness with a particular partner after some initial cooperative actions of the partner. By contrast, calculative reciprocity (based on fairness principles) continuously keeps track of its interaction balance with alters and adjusts its cooperativeness accordingly. Using an ecological simulation model, Back and Flache (2006) extended the de Vos model introducing variation in the extent to which a strategy follows commitment or calculative reciprocity behavior. This study showed that “strong” commitment strategies outperform “weaker” forms of commitment and various versions of calculative reciprocators under a wide range of conditions. However, a remaining major limitation of these analyses is that the spontaneous emergence of more sophisticated strategies was not considered. In particular, it was precluded that sophisticated cheaters emerge who optimally take advantage of the cooperativeness of commitment. Whether and to what extent this may be possible is crucial for
56
Chapter 3. The Evolutionary Advantage of Commitment
the validity of an explanation of interpersonal commitment behavior in terms of its evolutionary advantages in the human ancestral environment. Accordingly, in the present paper we provide a better test of evolutionary explanations for commitment by extending previous analyses with random mutation of strategies. In Section 3.2, we present our computational model and formulate conjectures. Section 3.3 contains results of simulation experiments, followed by a discussion and conclusions in Section 3.4.
3.2
Model
We use an abstract decision situation that we call the Delayed Exchange Dilemma (see de Vos et al., 2001; Back and Flache, 2006), or DED for short. The DED builds on the well-known repeated Prisoner’s Dilemma but contains two major extensions. First, it puts the problem of cooperation into a sequential exchange perspective, which is essentially a generalization of simultaneous exchange. Second, and more important is that it presents agents with a dilemma to choose interaction partners (see also e.g. Hayashi and Yamagishi, 1998). With these extensions the DED becomes ideal for studying commitment-related behavior in uncertain environments. The DED is played by n agents in successive rounds. Initially, all agents are endowed with fi points. In the beginning of each round Nature strikes a number of agents, each with a given individually independent probability Pd , who become in need of help from other agents. Agents who are struck by Nature are the initiators of interactions. Each of them asks another agent for help which is either provided or not. Providing help costs fh points. Moreover, help giving is time-consuming. Each agent can only provide help once during one round and only agents who are not distressed themselves may provide help. If a help request is turned down, the distressed agent may ask another agent for help but not more than m agents altogether within the same round, due to time restrictions. If an agent does not get help before the end of the round, it experiences fd loss in points. If the points of an agent fall below a critical threshold fc , the agent dies.
3.2.1
Modeling strategies
To explicitly study the evolutionary viability of commitment and fair reciprocity, we model preferences as a combination of commitment-related traits, fairness-related traits and a general cooperativeness trait. These traits determine the extent to which agents base their decisions on commitment- or fairness-related aspects of a decision situation. Equipped with these preferences agents decide about cooperation and also about choosing interaction partners.
3.2. Model
57
In particular, agents may face two different types of decision situations repeatedly in the DED. When they are hit by distress, they have to select an interaction partner to ask help from. On the other hand, when they themselves are asked to provide help they need to decide whether to provide it and in case of multiple requests whom to provide it to. In both cases, agents order possible interaction partners according to the attractiveness of interacting with them. Attractiveness is based on the individual preferences agents have with regard to past interaction histories. The attractiveness of agent aj for giving help to, calculated by agent ai , is formalized as: G G Uij = commG i · IN T F REQij + f airi · IN T BALij + coopi , G where commG i is the preference for commitment in giving, f airi is the preference for fairness in giving, and coopi is the preference for general cooperativeness. IN T F REQij is the proportion of cooperative interactions2 ai had with aj compared to the total number of cooperative interactions ai had. A cooperative interaction is defined as an interaction in which either ai helped aj or ai received help from aj . IN T BALij is the standardized interaction balance between agents ai and aj . To obtain this measure, we took into account both the balance of helps and the balance of refusals. The reason is that neither help balance nor refusal balance alone is sufficient to guarantee an overall balance in the exchange relationship. For example, suppose ai helped aj equally often as aj helped ai but ai refused to help aj ten times more often than aj refused to help ai . Despite the equal amount of help given, this exchange relationship clearly cannot be considered perfectly in balance. Technically, we calculated the measure as follows. We subtracted from 1.0 a measure of the overall standardized imbalance. The overall standardized imbalance is obtained by adding the difference of the number of times ai received help from aj and ai gave help to aj , and the difference of the number of times ai refused to give help to aj and aj refused to give help to ai , and dividing this by the total number of interactions they had. When comparing IN T F REQ and IN T BAL, notice that while a committed agent ai will find an interaction partner aj more attractive the more often it helped aj , a fair agent will be negatively influenced by the same fact. Note also that for simplicity, this model treats the impact of helping and refusing to help on the interaction balance as equally large. In the actual implementation, every time an agent has to make a decision, there is also a probability Pe that the agent will not use the above utility calculation but will choose randomly from the set of available decisions, each 2 Interactions take place always between exactly two agents. Possible interactions are giving help (cooperation) and refusing to help (defection). Asking for help is always followed by one of these.
58
Chapter 3. The Evolutionary Advantage of Commitment
being equally likely. This random error models noise in communication, misperception of the situation or simply miscalculation of the utility by the agent. Taking this random error into account increases the robustness of our results to noise. The attractiveness of agent aj for asking help from is defined in a similar way, the difference is that agents may put different weights on the two historyspecific decision parameters, and that there is no cooperativeness parameter: A A Uij = commA i · IN T F REQij + f airi · IN T BALij ,
Before agents make a decision, be it help giving or help asking, they calculate the corresponding one of these two types of attractiveness respectively for each agent who asked for help (U G ), or for each other agent in the population (U A ). In case of help giving, they choose a partner with the highest attractiveness, if that attractiveness is above an agent-specific threshold uti . Notice that IN T F REQij and IN T BALij are always smaller than or equal to 1. We allow commi , f airi and coopi to take values from [−1; 1]. Thus we allow the attractiveness threshold uti to take values from [−3; 3]. If the attractiveness of all possible agents is below the threshold attractiveness, no help is given to anyone. Otherwise, if there is more than one other agent with highest attractiveness3 , the agent selects one of the others with equal probability. In the case of help seeking, agents also choose a partner with the highest attractiveness but there is no threshold, i.e. agents in distress always ask someone for help. Definition 7 (Strategy). A strategy is a combination of four traits for help giving behavior (commG , f airG , coop, ut ) and two traits for help asking behavior (commA , f airA ).
3.2.2
Evolutionary dynamic
The heart of our model is an evolutionary dynamic that captures the random mutation of strategies and selection of objectively successful ones. The implementation of this process is based on the replicator dynamics (Taylor and Jonker, 1978). Broadly, the replicator dynamics dictates that if a generation of genotypes (strategies) undergoes reproduction, the net reproduction rate of a genotype is proportional to its relative success compared to other genotypes in the current generation. Genotypes that perform below average, in particular, have a negative reproduction rate. In our case, genotypes (strategies) represent subjective preferences. To prevent a population from growing without bounds, thus modeling resource scarcity in an implicit way, we keep the size of the population constant, 3 This
is unlikely, as the preference parameters are high precision real values and interaction histories tend to differ with time.
3.3. Conjectures
59
in the following way. At the end of each round we count how many agents have died and replace them with new agents in the next round. Each new agent A has the same strategy as a randomly selected other agent B, present in the population who has reached a minimum age n (measured in the number of interactions it had). The probability of choosing this other agent B is proportionate to the share of points B holds within the group of all agents older than n. Before A is added to the population, with probability Pmut , its strategy may undergo mutation. A mutation occurs in exactly one, randomly chosen trait, with equal probabilities for all traits, thus P = 91 for each trait. The new value of the trait is a uniformly distributed random value from the interval [−3; 3] for the attractiveness threshold, and from [−1; 1] for all other traits.
3.3
Conjectures
To guide the simulation experiments, in the following we formulate a number of conjectures derived from previous work. Definition 8 (Stability). The stability of a strategy s is equal to the number of consecutive rounds that it existed in a population in a given simulation run, counting from the first round it appeared until the round in which it became extinct. A strategy s is infinitely stable if it does not become extinct. Generalizing from analytical results about the evolutionary stability of strategies in repeated games that are simpler than the DED (cf. Bendor and Swistak, 2001), we expect that there is no single strategy that is superior to all others in the dilemma we study. In other words, for every incumbent strategy there exists another (mutant) strategy that can take advantage of the incumbent’s weakness. Conjecture 3.1. There is no infinitely stable strategy in an infinitely played game of DED. Nevertheless, the length of time a strategy exists (stability) carries an important message about its viability. Since mutations constantly arise and threaten to push other strategies out of the population, stability is an indicator for the number of attacks a strategy can withstand. Therefore, the stability of a strategy will be one of the indicators of its viability4 . The other measure is typical longevity within a strategy (variable longevity, the average age at death of agents belonging to a strategy). Note that in our model, there is no 4 We will not use here stability concepts from the evolutionary game theory literature (e.g. evolutionary stability or asymptotic stability) because they do not allow the expression of the relative stability of strategies.
60
Chapter 3. The Evolutionary Advantage of Commitment
upper age limit on reproducibility, in other words, agents keep reproducing until they die, which makes longevity a suitable measure for viability. Back and Flache (2006) found that the most successful strategies in the DED exhibited some level of interpersonal commitment and that committed agents outcompeted fair reciprocators. These results suggest the following two conjectures, which we will test under the new assumption of random mutation: Conjecture 3.2. Individual preferences for interpersonal commitment and fairness have a positive effect on viability. Conjecture 3.3. The positive effect of commitment preferences on viability is stronger than the effect of fairness preferences. According to de Vos et al. (2001), commitment works best under harsh conditions: the more agents are challenged by Nature to survive, the more compelled they are to cooperate with each other in durable relationships. More technically, they found that the larger the probability of distress, the larger the proportion of commitment strategies surviving, relative to the defector strategy. This leads us to test: Conjecture 3.4. Environmental harshness has a positive interaction effect on stability with the level of cooperation and interpersonal commitment of a strategy.
3.4
Results
Binmore (1998) argued forcefully that the outcome of computer tournaments and simulations of evolutionary dynamics strongly depends on the set of strategies that are initially present in a population. To avoid our results becoming biased by a restrictive set of starting conditions, we ran several hundred replications of our simulation runs, each time with a population whose initial strategy is randomly chosen from the strategy space defined by the six traits. We did not find any significant effects of features of initial starting strategies on the outcomes of simulation runs. The reason is that soon after the initial rounds of a simulation run, mutation ensures the emergence of a large variety of different strategies in the population. We allow this population to play the DED game. In the course of the game agents start to lose points, some of them eventually die, while others reproduce. At some point, random mutations occur in the initial strategy, creating a potential invader. The better a mutated strategy performs in the DED compared to agents of the original strategy, the larger is its probability of reproducing and increasing its proportion within the agent population. The simulation run ends with either the extinction of all agents5 or after an arbitrarily chosen 5 Extinction
is possible if all agents die within one round and thus there is no basis for the distribution of strategies in the next generation.
3.4. Results
61
large number of rounds (10 million). We then repeat the simulation run with another, randomly generated initial population. During each simulation run we record all strategies and their key characteristics that have ever appeared through random mutations. These characteristics include on the individual-level the traits of the strategy (commA , f airA , coop, ut , commG , f airG ); the average longevity measured in rounds of game play on the strategy-level; and finally a population-level variable measuring the overall level of cooperation and defection (SOCCOOP)6 .
3.4.1
Initial parameters
To preserve comparability of our results, we started our simulations with the same initial parameters (where applicable) that were used in earlier work. These are Pd = 0.2, fh = 1, fd = 20, fi = 100, fc = 0, N = 25, Pe = 0.05, m = 2. (For the meaning of each parameter, please consult the Model section above.) These parameters impose a set of conditions under which for strictly instrumental agents the choice between purposeful defection and (conditional) cooperation is as difficult as possible. The parameters are determined such that in a simplified two-person version of the game, perfectly rational actors would be indifferent between choosing for a conditionally cooperative and a fully defecting strategy if they meet a conditionally cooperative partner. In this way, we implement a setting in which the problem of cooperation is particularly hard to solve and provide thus a strict test of the viability of cooperative strategies, including commitment and fairness. If cooperation placed an excessively high burden on agents, or conversely, if cooperation entailed no real sacrifices, the model would hardly yield any interesting insights. (For the detailed game theoretical derivation using trigger strategies, see A.) We refer to this parameter setting as the baseline condition. After obtaining results for the baseline condition, we conduct experiments in which we systematically vary the level of environmental harshness (fd ). Furthermore, we run additional experiments with varying parameter combinations to test the sensitivity of results to variation in model parameters.
3.4.2
Stability
In support of conjecture 3.1, our simulation results show that strategies change endlessly in all initial parameter settings – we found no infinitely stable strategy in the DED. We simulated 175 runs altogether, each of which started with a different randomly chosen initial strategy and consisted of maximally 10,000,000 rounds. During these runs more than 4.7 million mutations took 6 SOCCOOP
measures the difference between the per-round average number of cooperation (helps) and defection (refusals) in the entire population.
62
Chapter 3. The Evolutionary Advantage of Commitment
place altogether, generating as many strategies. However, in none of these runs have we recorded any strategy that existed longer than 220,000 rounds. We may of course simply have not encountered the infinitely stable strategies during our random walks in this vast strategy space. However, judging by the vast coverage of the strategy space by our method, this seems implausible. A plausible explanation for the lack of infinitely stable strategies is that for each strategy there exists a better response that takes advantage of the strategy’s weakness. Sooner or later mutations generate this better response and the original strategy is gradually pushed out of existence. If a strategy is too cooperative, opportunistic exploiters take advantage of this and flourish. Later, in a harsh world of mainly exploiters, where everybody is suffering, two cooperators who appear randomly at the same time and find each other will survive and reproduce more easily than others, given that they have a sound method of excluding defectors from cooperative interactions. That cooperativeness eventually declines again may be explained by the gradual loss of the ability to exclude defectors through “evolutionary drift” (e.g. Bendor and Swistak, 2001) or by the emergence of new defecting mutants who have developed the ability to behave such that they are not excluded from exchanges between incumbent cooperators. Figure 3.1 illustrates these dynamics of average age at death and helping behavior, for a typical simulation run. The upper part of the figure shows how the average age at death (measured in interactions) changes over time within one simulation round. Compare this figure with the level of cooperation, generated for the same simulation run, in the lower part of the figure: periods of high refusal rates coincide with short lives.
3.4.3
The importance of interpersonal commitment
Conjectures 3.2 and 3.3 relate the strength of the commitment preference within a strategy directly to viability, the average length of an agent’s life within a strategy. Commitment is measured by the commG and commA traits, distinguished for giving and asking respectively. The higher these traits are, within the [-1;1] interval, the more an agent is inclined to choose and cooperate with long-term interaction partners. If they are positive, the agent has a preference for commitment; if they are negative, the agent has a preference against being committed; and when the values are close to zero, the agent is indifferent to the concept of commitment. What we find is that among the most stable 1,355 strategies (where stability is at least 50,000 rounds), 776 strategies (57.3%) are positive on both commitment traits. Among the same strategies, only 385 (28.4%) are positive on both f airG and f airA , and 717 (52.9%) on coop. This suggests that if a strategy is highly stable, its decision-making process is likely to be guided by prefer-
3.4. Results
63
ences for unconditional cooperation with old interaction partners. These preferences appear to be far more important for success than being fair or simply being cooperative (coop trait). To get a closer insight into the separate contributions of the traits to a strategy’s success and to compare in particular the relative importance of the f airG and f airA traits to the importance of the commG and commA traits, we conducted a linear regression analysis with average longevity within a strategy as the dependent variable (see Table 3.1). Before performing the analysis we filtered out highly unstable strategies (STAB