Agent_Zero: Toward Neurocognitive Foundations for Generative Social Science

Agent_Zero

Princeton Studies in Complexity

SERIES EDITORS

Simon A. Levin (Princeton University)

Steven H. Strogatz (Cornell University)

Lars-Erik Cederman, Emergent Actors in World Politics:

How States and Nations Develop and Dissolve

Robert Axelrod, The Complexity of Cooperation:

Agent-Based Models of Competition and Collaboration

Peter S. Albin, Barriers and Bounds to Rationality:

Essays on Economic Complexity and Dynamics in Interactive Systems. Edited and with an introduction by Duncan K. Foley

Duncan J. Watts, Small Worlds:

The Dynamics of Networks between Order and Randomness

Scott Camazine, Jean-Louis Deneubourg, Nigel R. Franks, James Sneyd, Guy Theraulaz, and Eric Bonabeau, Self-Organization in Biological Systems

Peter Turchin, Historical Dynamics: Why States Rise and Fall

Andreas Wagner, Robustness and Evolvability in Living Systems

Mark Newman, Albert-Laszlo Barabasi, and Duncan Watts, eds., The Structure and Dynamics of Networks

J. Stephen Lansing, Perfect Order: Recognizing Complexity in Bali

Joshua M. Epstein, Generative Social Science:

Studies in Agent-Based Computational Modeling

John H. Miller and Scott E. Page, Complex Adaptive Systems:

An Introduction to Computational Models of Social Life

Michael Laver and Ernest Sergenti, Party Competition: An Agent-Based Model

Joshua M. Epstein, Agent_Zero:

Toward Neurocognitive Foundations for Generative Social Science

Agent_Zero

Toward Neurocognitive Foundations for Generative Social Science

Joshua M. Epstein

Princeton University Press

Princeton and Oxford

Copyright © 2013 by Princeton University Press

Published by Princeton University Press, 41 William Street, Princeton, New Jersey 08540

In the United Kingdom: Princeton University Press, 6 Oxford Street, Woodstock, Oxfordshire OX20 1TW

press.princeton.edu

All Rights Reserved

Library of Congress Cataloging-in-Publication Data

Epstein, Joshua M., 1951–

Agent zero : toward neurocognitive foundations for generative social science / Joshua M. Epstein.

pages cm. — (Princeton studies in complexity)

Summary: The Final Volume of the Groundbreaking Trilogy on Agent-Based Modeling In this pioneering synthesis, Joshua Epstein introduces a new theoretical entity: Agent Zero. This software individual, or agent, is endowed with distinct emotional/affective, cognitive/deliberative, and social modules. Grounded in contemporary neuroscience, these internal components interact to generate observed, often far-from-rational, individual behavior. When multiple agents of this new type move and interact spatially, they collectively generate an astonishing range of dynamics spanning the fields of social conflict, psychology, public health, law, network science, and economics. Epstein weaves a computational tapestry with threads from Plato, Hume, Darwin, Pavlov, Smith, Tolstoy, Marx, James, and Dostoevsky, among others. This transformative synthesis of social philosophy, cognitive neuroscience, and agent-based modeling will fascinate scholars and students of every stripe. Epstein’s computer programs are provided in the book or on its Princeton University Press website, along with movies of his computational parables. Agent Zero is a signal departure in what it includes (e.g., a new synthesis of neurally grounded internal modules), what it eschews (e.g., standard behavioral imitation), the phenomena it generates (from genocide to financial panic), and the modeling arsenal it offers the scientific community. For generative social science, Agent Zero presents a groundbreaking vision and the tools to realize it—Provided by publisher.

Includes bibliographical references and index.

ISBN 978-0-691-15888-4 (hardback)

1. Social sciences—Computer simulation. 2. Social sciences—Mathematical models. I. Title.

H61.3.E6697 2013

300.1—dc23         2013018009

British Library Cataloging-in-Publication Data is available

The research for this book was funded by a 2008 NIH Director’s Pioneer Award

This book has been composed in Minion Pro and Fairfield

Printed on acid-free paper. ∞

Printed in the United States of America

1 3 5 7 9 10 8 6 4 2

For Matilda,

My lark at break of day arising …

Contents

Foreword

I SEE THIS BOOK as the third in a trilogy on generative social science.

VOLUME I

The first volume of the trilogy was Growing Artificial Societies: Social Science from the Bottom Up (MIT Press/Brookings Press), with coauthor Robert Axtell. Published in 1996, this introduced the Sugarscape agent-based model, and the notion of a generative explanation of social phenomena. Sugarscape was a single sweeping exploratory artificial society, with glimmerings of a mature generative epistemology.

VOLUME II

For the subsequent decade, with diverse colleagues, I applied agent-based modeling to a broad spectrum of fields—economics, archaeology, conflict, epidemiology, spatial games, and the dynamics of norms—and thought more deeply about the philosophy of agent-based social science. The results are collected in the second volume of the trilogy: Generative Social Science: Studies in Agent-Based Computational Modeling, published in 2006 by Princeton University Press. Relative to Growing Artificial Societies, Generative Social Science presented a collection of more tightly focused and specifically explanatory exercises, and a far more extended and mature generative epistemology.

VOLUME III

Generative Social Science ended with a challenge: Grow Raskolnikov! By this, I meant agents with more fully developed—and so conflicted—inner lives.¹ In effect, it was a call for greater cognitive realism. The present book is my response. It introduces a new theoretical entity, Agent_Zero, whose observable behavior is itself generated by the interaction, indeed conflict, of affective, deliberative, and social components. Passion, reason, and social forces, in other words, are all at play in Agent_Zero’s observable behaviors, which span a wide array of fields, including economics, health, conflict, social psychology, and endogenous network dynamics. Needless to say, the relationship between passion and reason within the individual, and the relation of the individual to society, are among the more enduring questions in philosophy. And while I do not claim to resolve them, I do claim to treat them in a new way. A significant volume of contemporary cognitive neuroscience is employed in constructing Agent_Zero, whom I offer as a new, neurocognitively grounded, foundation for generative social science—hence the subtitle of this book. However, as is repeated numerous times, I use only selected neuroscience in developing this particular agent model. I do not purport to encompass—much less to advance—any area of neuroscience itself. Indeed, through the good offices of a number of very fine neuroscientists, I may have avoided insulting their fascinating and fast-moving discipline.

¹Indeed, the name Raskolnikov is from the Russian, raskólʾnik, meaning, roughly, schismatic.

Preface

EMOTIONAL, COGNITIVE, AND SOCIAL FACTORS shape the behavior of individuals in groups and hence shape the emergence of important social dynamics, from genocide to financial panic. I wish to generate such social dynamics from the bottom up, in social networks of neurocognitively plausible individuals. To this end, I introduce a new theoretical entity, Agent_Zero, endowed with interacting emotional/affective, cognitive/deliberative, and social modules. Agent_Zero’s affective component is based on the Rescorla-Wagner model of conditioning and extinction, supported by recent science on the neural mechanisms of fear conditioning specifically. The agent’s cognitive (deliberative) component reflects well-documented biases and heuristics in the estimation of probabilities (e.g., sample selection bias). Agents belong to social networks, and the social component exhibits contagion effects. But, crucially, it is not observable behavior that is transmitted in this model, but disposition. I define this here as an explicit function of (a) the individual’s emotion and cognition (passion and reason), and (b) others’ affective and deliberative states. Action is binary and is triggered when individual disposition exceeds threshold. These thresholds, and susceptibility to dispositional contagion, can be heterogeneous across agents, all of whom can exhibit emotional inertia and memory of events. The same basic model, interpreted and extended variously, is shown to generate core phenomena in the fields of social conflict, psychology, public health, law, network dynamics, and economics. Mathematical and spatial agent-based computational versions of the general model are presented.

I believe Agent_Zero to be a departure in what it includes (e.g., neurally grounded internal modules), what it excludes (e.g., standard behavioral imitation), the range of phenomena it generates, and the set of tools it offers the field. Overall, I submit Agent_Zero as a step toward unified—and neurocognitively grounded—foundations for generative social science.

Acknowledgments

THIS RESEARCH WAS SUPPORTED entirely by a National Institutes of Health Director’s Pioneer Award (DP1). This afforded me unparalleled scientific freedom and resources to found a new Center for Advanced Modeling in the Social, Behavioral, and Health Sciences at Johns Hopkins University, where this research was primarily conducted. For these exceptional opportunities, I am eternally grateful to the NIH. The Center would not have been possible without the vision and support of Dr. Gabor Kelen, Chairman of the Department of Emergency Medicine, to whom I am profoundly grateful. I also wish to thank my colleagues in the department for the climate of uncompromising excellence in which it is my privilege to work. My closest colleague in this project has been Julia Chelen, whom I thank for superb research assistance and invaluable consultations throughout. For detailed reviews of the manuscript, I am indebted to Robert Axelrod, Jon Parker, and Lisa Feldman-Barrett. For important reactions, from a variety of scientific perspectives, I also thank Paul Smaldino, Eili Klein, Michael Makowsky, David Broniatowski, and Erez Hatna. I am grateful for numerous sessions on learning theory with Peter Holland. For many formative discussions and for cohosting a two-day NIH Workshop on the model, I thank John Cacioppo. For their many insights and comments, I am grateful to the workshop participants: Colin Camerer, Julia Chelen, William Eaton, Stephen Eubank, Thomas Glass, Joseph Harrington, Melissa Healy, Peter Holland, Gabor Kelen, Ethan Kross, Jon Parker, Elizabeth Phelps, Paul Torrens, and Richard Rothman. I am deeply indebted to the Princeton University Press external reviewers Duncan Foley, Paul Slovic, and John Steinbruner for thorough readings, penetrating insights, and valuable suggestions for improvement. I also thank the book’s Princeton University Press editorial and production team for their engagement and expertise: Vickie Kearn, Mark Bellis, Dimitri Karetnikov, and Linda Thompson. I thank Donald Burke for many stimulating colloquies about the possibility of a unified behavioral health science. The first year of this research was conducted at The Brookings Institution, which is gratefully acknowledged. Ross Hammond, Kislaya Prasad, and Peyton Young offered valuable comments during this phase of the research. For countless discussions of generative minimalism and the philosophy of science, I thank my brother, Samuel David Epstein. Finally, for their love and support, I thank my wife, Melissa Healy, our son, Joey, and our daughter, Matilda, to whom this book is dedicated, in memory of my parents, Joseph and Lucille Epstein.

Agent_Zero

INTRODUCTION

MOTIVATION

IN HIS TREATISE OF HUMAN NATURE, David Hume (1739; 2000 ed.) famously wrote, Reason is … the slave of the passions.² In using the term slave, however, Hume’s point is not that the passions always prevail over reason in a tug of war,³ but that the two are, in some sense, incommensurable. Distinguishing the passions from factual (true/false) claims, he writes, Tis impossible [that] they [the passions] can be pronounced either true or false, and be either contrary or conformable to reason (p. 458). I take this nonconformability to mean that, as a modeling proposition, passion and reason—emotion and ratiocination, if you prefer—belong on different axes. They might be seen as basis elements of a dispositional vector space. If we adopt Hume’s geometry, then beyond passion and reason there is a third basis element, or axis, inspired by Spinoza, who wrote that Man is a social animal. ⁴ In this view, then, our actions may be motivated by the passions, but are influenced by reason—and by society.

Needless to say, Hume was not the last, nor Spinoza the first, to philosophize on the roles of emotion, deliberation, and social influence in human behavior. Indeed, innumerable contrasting perspectives on the balance among them have been articulated over the centuries. I do not purport to end this dialogue. I do claim, however, to inject into that discussion a new approach.

Specifically, my central objective here is to develop a simple explicit model of individual behavior in groups that includes some representation of the passions, of (imperfect) reason, and of social influence. In other words, I will offer an exploratory synthesis of three (partially understood and obviously intertwined) processes:

•  The emotional

•  The cognitive

•  The social

To this end, I introduce a new theoretical entity, Agent_Zero, endowed with emotional/affective, cognitive/deliberative, and social modules whose—often nonconscious—interactions determine his or her observed behavior.

Generate Social Dynamics

The aim is to generate recognizable dynamics of social importance. Specifically, we humans often do things in groups that we would not do alone (Le Bon, 1895; Canetti, 1984; Mackay, 1841, Browning, 1998). We do things for which we have no basis in evidence. Indeed, we do them knowing we have no evidence; and sometimes, despite this, we are even the first in the group to do them!

Dispositional Contagion, Not Behavioral Imitation

Notice that in this latter case, the imitation of behavior cannot be the mechanism. Clearly, if I am the first actor, I cannot be imitating anyone’s behavior, because no one before me has acted! But, more to the point, in this model my successors are not imitating my behavior either. No agent is imitating any other’s behavior as defined in this model. Here, behavior is binary—agents act or do not—and, by design, this Boolean variable does not appear as an independent (right-hand side) variable in any agent’s algorithm (discussed more fully later). Of course, agents may end up doing identical things, but the mechanism introduced here is dispositional contagion, not monkey see, monkey do imitation of observable binary action, which, as I say, is barred mathematically. Without positing behavioral imitation in the usual sense,⁵ the Agent_Zero model will be shown to generate a wide range of social dynamics.

A Core Target

As a core target for the model, I want to grow the person who feels no aversion to black people, who has never had any direct evidence or experience of black wrongdoing (his empirical estimate of this probability is and always has been literally zero), and who yet initiates the lynching. Though, as first mover, he does not copy the behavior of others, a type of contagion is at play, but it is dispositional. He is, of course, an ideal type, but an important one.⁶

Mathematical social science has largely attempted to characterize the rational; I am interested in generating behavior that is far from rational. And I believe there is more at play than mere bounds on information or computing resources. While cases of collective cruelty abound, there are also examples of collective resistance to it. A general theory should generate both. I will attempt to do so here.

Specifically, Agent_Zero’s observable behavior will result from the interplay of (a) emotional/affective, (b) cognitive/deliberative, and (c) network/contagion components. As defined here, the agent’s total disposition to act will be an explicit mathematical function of these. This disposition will be a real number. If it exceeds some threshold, Agent_Zero acts. Otherwise she does not.

Each model component (affective, deliberative), of course, is an entire discipline in its own right. And the crude models set forth will doubtless merit criticism from domain experts. But I am less interested in the accuracy of the components than in the generative capacity of the synthesis. The components can—indeed must—be refined. But, as far as I know, this particular synthesis has not been attempted. So, to me, it is less important to get the components finished than to get the synthesis started, as simply and understandably as possible.

While it impinges on a number of enduring philosophical issues, this synthesis aims to fill an important gap in contemporary social science. In statistical social science, relationships among aggregate macroscopic variables are assessed econometrically. The individual, as such, is not represented. In mathematical economics and game theory, the individual is represented, but the representation is not grounded in cognitive neuroscience.⁷ But laboratory neuroscience—transformed by advances in imaging—is focused on single individuals (e.g., fMRI subjects in controlled laboratory experiments), not on networks of individuals.⁸ So, what happens in networks of cognitively plausible individuals? Indeed, by what mechanisms do cognitively plausible individuals generate networks? This book explores these questions mathematically, and computationally—in the latter case, by the method of agent-based modeling.

Generative Minimalism

This effort is exploratory and unapologetically theoretical. It is an exercise in generative minimalism. I do not fit the basic model or its extensions to numerical data, though this is an obvious candidate for future research, nor do I claim to predict events.⁹ Rather, I hope to demonstrate that the Agent_Zero model is sufficient to generate core qualitative social dynamics across a wide range of domains. These demonstrations establish the model’s generative explanatory candidacy, in the language of J. M. Epstein (2006). For a full discussion of the generative explanatory epistemology, generative minimalism, and examples of empirically calibrated agent models, see J. M. Epstein (2006).¹⁰

Not Modeling Brain Regions

There is another, extremely important, qualification. We will, of course, have occasion to discuss a number of brain regions: the amygdala, hippocampus, and prefrontal cortex, for example. But I am emphatically not modeling brain regions. To clinicians, brain regions are of interest because functional localization facilitates therapy (e.g., neurosurgery). But to social theorists brain tissue is interesting exactly and only in so far as it licenses model design or model interpretation.

What is plausible for humans and what is not? The neuroscience will both illuminate and explain those limits, limits that belong in the model and that have been largely ignored by social science. The brain discussions that follow are meant only to support that modeling aim. If my account of the tissue science is wanting (as it surely is) or if the state of tissue science advances (as it surely will), I do not believe that per se endangers the mathematical model. One wants to develop a mathematical/computational framework that can exploit and accommodate the evolving tissue science but is not hostage to today’s snapshot of it, for surely there are few fields developing faster. At the same time, one’s model should also be sufficiently binding as to admit falsification. I hope to strike this difficult balance.¹¹

Thus, I will display the formal models I have in mind and try to under-gird them with what little cognitive neuroscience science I know, hoping that the attempt will elicit the empathetic interest and collaboration of domain experts. Now to the model’s formal constituents.

THE MODEL COMPONENTS

For the affective/emotional component, we will use the essentially Pavlovian theory of associative learning; specifically, a generalization of the Rescorla-Wagner (1972) model of conditioning. For the cognitive—ratiocinative and deliberate—component, agents will form a probability based on local sampling (which can introduce bias). For the social component, I will use a network transmission model, but with a crucial wrinkle. Most work on contagion in networks (including some of my own) posits the conscious observation and imitation of behavior. Very importantly, however, in the present model it is not observable (binary) behavior that is contagious in these networks, but dispositions. Finally, as a starting point, the function combining these affective, deliberative,