Book Search:  


Google full text of our books:


Generative Social Science:
Studies in Agent-Based Computational Modeling
Joshua M. Epstein

Book Description | Endorsements | Table of Contents

COPYRIGHT NOTICE: Published by Princeton University Press and copyrighted, © 2007, by Princeton University Press. All rights reserved. No part of this book may be reproduced in any form by any electronic or mechanical means (including photocopying, recording, or information storage and retrieval) without permission in writing from the publisher, except for reading and browsing via the World Wide Web. Users are not permitted to mount this file on any network servers. Follow links for Class Use and other Permissions. For more information, send e-mail to

This file is also available in Adobe Acrobat PDF format

Chapter 1



This article argues that the agent-based computational model permitsa distinctive approach to social science for which the term “generative” is suitable. In defending this terminology, features distinguishing the approach from both “inductive” and “deductive” science are given. Then, the following specific contributions to social science are discussed: The agent-based computational model is a new tool for empirical research.It offersa natural environmentforthestudyof connectionist phenomena in social science. Agent-based modeling provides a powerful way to address certain enduring—and especially interdisciplinary—questions. It allows one to subject certain core theories—such as neoclassical microeconomics—to important types of stress (e.g., the effect of evolving preferences). It permits one to study how rules of individual behavior give rise—or “map up”—to macroscopic regularities and organizations. In turn, one can employ laboratory behavioral research findings to select among competing agent-based (“bottom up”) models. The agent-based approach may well have the important effect of decoupling individual rationality from macroscopic equilibrium and of separating decision science from social science more generally. Agent-based modeling offers powerful new forms of hybrid theoretical-computational work; these are particularly relevant to the study of non-equilibrium systems. The agent-based approach invites the interpretation of society as a distributed computational device, and in turn the interpretation of social dynamics as a type of computation. This interpretation raises important foundational issues in social science—some related to intractability, and some to undecidability proper. Finally, since “emergence” figures prominently in this literature, I take up the connection between agent-based modeling and classical emergentism, criticizing the latter and arguing that the two are incompatible.

Generative Social Science

The agent-based computational model—or artificial society—is a new scientific instrument.1 It can powerfully advance a distinctive approach to social science, one for which the term “generative” seems appropriate. Iwill discuss this term more fully below, but in a strong form, the central ideais this:Tothe generativist, explainingthe emergence2of macroscopic societal regularities, such as norms or price equilibria, requires that one answerthe following question:

The Generativist’s Question

*How could the decentralized local interactions of heterogeneous autonomous agents generate the given regularity?

The agent-based computational model is well-suited to the study of this question since the following features are characteristic:3


Representative agent methods—common in macroeconomics—are not used in agent-based models (see Kirman 1992). Nor are agents aggregated into a few homogeneous pools. Rather, agent populations are heterogeneous; individuals may differ in myriad ways—genetically, culturally, by social network, by preferences—all of which may change or adapt endogenously over time.


There is no central, or “top-down,” control over individual behavior in agent-based models. Of course, there will generally be feedback from macrostructures to microstructures, as where newborn agents are conditioned by social norms or institutions that have taken shape endogenously through earlier agent interactions. In this sense, micro and macro will typically co-evolve. But as a matter of model specification, no central controllers or other higher authorities are posited ab initio.


Events typically transpire on an explicit space, which may be a landscape of renewable resources, as in Epstein and Axtell (1996), an n-dimensional lattice, or a dynamic social network. The main desideratum is that the notion of “local” be well posed.


Typically, agents interact with neighborsinthisspace (and perhaps with environmental sites in their vicinity). Uniform mixing is generically not the rule.4 It is worth noting that although this next feature is logically distinct from generativity, many computational agent-based models also assume:


There are two components of this: bounded information and bounded computing power. Agents do not have global information, and they do not have infinite computational power. Typically, they make use of simple rules based on local information (see Simon 1982 and Rubinstein 1998).

The agent-based model, then, is especially powerful in representing spatially distributed systems of heterogeneous autonomous actors with bounded information and computing capacity who interact locally.

The Generativist’s Experiment

In turn, given some macroscopic explanandum—a regularity to be explained—the canonical agent-based experiment is as follows:

** Situate an initial population of autonomous heterogeneous agents in a relevant spatial environment; allow them to interact according to simple local rules, and thereby generate—or “grow”—the macroscopic regularity from the bottom up.5

Concisely, ** is the way generative social scientists answer *.In fact, this type of experiment is not new6 and, in principle, it does not necessarily involve computers.7 However, recent advances in computing, and the advent of large-scale agent-based computational modeling, permit a generative research program to be pursued with unprecedented scope and vigor.


A range of important social phenomena have been generated in agent-based computational models, including: right-skewed wealth distributions (Epstein and Axtell 1996), right-skewed firm size and growth rate distributions (Axtell 1999), price distributions (Bak et al. 1993), spatial settlement patterns (Dean et al. 1999), economic classes (Axtell et al. 2001), price equilibria in decentralized markets(Albin and Foley 1990; Epstein and Axtell 1996), trade networks (Tesfatsion 1995; Epstein and Axtell 1996), spatial unemployment patterns(Topa 1997), excess volatility in returns to capital (Bullard and Duffy 1998), military tactics (Ilachinski 1997), organizational behaviors (Prietula, Carley, and Gasser 1998), epidemics (Epstein and Axtell 1996), traffic congestion patterns (Nagel and Rasmussen 1994), cultural patterns (Axelrod 1997c; Epstein and Axtell 1996), alliances (Axelrod and Bennett 1993; Cederman 1997), stock market price time series (Arthur et al. 1997), voting behaviors (Kollman, Miller,andPage 1992), cooperationinspatial games (Lindgren and Nordahl 1994; Epstein 1998; Huberman and Glance 1993; Nowak and May 1992; Miller 1996), and demographic histories (Dean et al. 1999). These examples manifest a wide range of (often implicit) objectives and levels of quantitative testing.

Before discussing specific models, it will be useful to identify certain changes in perspective that this approach may impose on the social sciences. Perhaps the most fundamental of these changes involves explanation itself.

Explanation and Generative Sufficiency

Agent-based models provide computational demonstrations that a given microspecification is in fact sufficient to generate a macrostructure of interest. Agent-based modelers may use statistics to gauge the generative sufficiency of a given microspecification—to test the agreement between real-world and generated macro structures. (On levels of agreement, see Axtell and Epstein 1994.) A good fit demonstrates that the target macrostructure—the explanandum—be it a wealth distribution, segregation pattern, price equilibrium, norm, or some other macrostructure, is effectively attainable under repeated application of agent-interaction rules: It is effectively computable by agent society. (The view of society as a distributed computational device is developed more fully below.) Indeed, this demonstration is taken as a necessary condition for explanation itself.To the generativist—concerned with formation dynamics— it does not suffice to establish that, if deposited in some macroconfiguration, the system will stay there. Rather, the generativist wants an account of the configuration’s attainment by a decentralized system of heterogeneous autonomous agents. Thus, the motto of generative social science, if you will, is: If you didn’t grow it, you didn’t explain its emergence. Or, in the notation of first-order logic:

(∀x)(¬Gx ⊃¬Ex) (1)

It must be emphasized that the motto applies only to that domain of problems involving the formation or emergence of macroscopic regularities. Proving that some configuration is a Nash equilibrium, for example, arguably does explain its persistence, but does not account for its attainment.8

Regarding the converse of expression (1), if a microspecification, m, generates a macrostructure of interest, then m is a candidate explanation. But it may be a relatively weak candidate; merely generating a macrostructure does not necessarily explain its formation particularly well. Perhaps Barnsley’s fern (Barnsley 1988) is a good mathematical example. The limit object indeed looks very much like a black spleenwort fern. But—under iteration of a certain affine function system—it assembles itself in a completely unbiological way, with the tip first, then a few outer branches, eventually a chunk of root, back to the tip, and so forth—not connectedly from the bottom up (now speaking literally).

It may happen that there are distinct microspecifications having equivalent generative power (their generated macrostructures fit the macro-data equally well). Then, as in any other science, one must do more work, figuring out which of the microspecifications is most tenable empirically. In the context of social science, this may dictate that competing microspecifications with equal generative power be adjudicated experimentally—perhaps in the psychology lab.

In summary, if the microspecification m does not generate the macrostructure x, then m is not a candidate explanation. If m does generatex,itisa candidate.9If thereis morethan one candidate, further work is required at the micro-level to determine which m is the most tenable explanation empirically.10

For most of the social sciences, it must be said, the problem of multiple competing generative accounts would be an embarrassment of riches. The immediate agenda is to produce generative accounts per se.The principal instrument in this research program is the agent-based computational model. And as the earlier examples suggest, the effort is underway.

This agenda imposes a constructivist (intuitionistic) philosophy on social science.11 In the air is a foundational debate on the nature of explanation reminiscent of the controversy on foundations of mathematics in the 1920s–30s. Central to that debate was the intuitionists’ rejection of nonconstructive existence proofs (see below): their insistence that meaningful “existence in mathematics coincides with constructibility” (FraenkelandBar-Hillel1958,207).Whilethespecificsareof course different here—and I am not discussing intuitionism in mathematics proper—this is the impulse, the spirit, of the agent-based modelers: If the distributed interactions of heterogeneous agents can’t generate it, then we haven’t explained its emergence.

Generative versusInductive and Deductive

From an epistemological standpoint, generative social science, while empirical (see below), is not inductive, at least as that term is typically used in the social sciences (e.g., as where one assembles macroeconomic data and estimates aggregate relations econometrically). (For a nice introductionto general problemsof induction, beginning with Hume,see Chalmers 1982. On inductive logic, see Skyrms 1986. For Bayesians and their critics, see, respectively, Howson and Urbach 1993 and Glymour 1980.)

The relation of generative social science to deduction is more subtle. The connection is of particular interest because there is an intellectual tradition in which we account an observation as explained precisely when we can deduce the proposition expressing that observation from other, more general, propositions. For example, we explain Galileo’s leaning Tower of Pisa observation (that heavy and light objects dropped from the same height hit the ground simultaneously) by strictly deducing, from Newton’s Second Law and the Law of Universal Gravitation, the following proposition: “The acceleration of a freely falling body near the surface of the earth is independent of its mass.” In the present connection, we seek to explain macroscopic social phenomena. And we are requiring that they be generated in an agent-based computational model. Surprisingly, in that event, we can legitimately claim that they are strictly deducible. In particular, if one accepts the Church-Turing thesis, then every computation—including every agent-based computation—can be executed by a suitable register machine (Hodel 1995; Jeffrey 1991). It is then a theorem of logic and computability that every program can be simulatedbya first-order language.In particular, withNdenoting the natural numbers:

Theorem. Let P be a program. There is a first-order language L, and for each a ∈ N a sentence C(a) of L, such that for all a ∈ N, the P-computation with input a halts ⇔ the sentence C(a)is logically valid.


This theorem allows one to use the recursive unsolvability of the halting problem to establish the recursive unsolvability of the validity problem in first-order logic (see Kleene 1967). Explicit constructions of the correspondence between register machine programs and the associated logical arguments are laid out in detail by Jeffrey (1991) and Hodel (1995). The point here is that for every computation, there is a corresponding logical deduction. (And this holds even when the computation involves “stochastic” features, since, on a computer, these are produced by deterministic pseudo-random number generation (see Knuth 1969). Even if one conducts a statistical analysis over some distribution of runs—using different random seeds—each run is itself a deduction. Indeed, it would be quite legitimate to speak, in that case, of a distribution of theorems.)12 In any case, from a technical standpoint, generative implies deductive, a point that will loom large later, when we argue that agent-based modeling and classical emergentism are incompatible.

Importantly, however, the converse does not apply: Not all deductive argument has the constructive character of agent-based modeling. Nonconstructive existence proofs are obvious examples. These work as follows: Suppose we wish to prove the existence of an x with some property (e.g., that it is an equilibrium). We take as an axiom the so-called Law of the Excluded Middle that (i) either x exists or x does not exist. Next, we (ii) assume that x does not exist, and (iii) derive a contradiction. From this we conclude that (iv) x must exist. But we have failed to exhibit x, or indicate any algorithm that would generate it, patently violating the generative motto (1).13 The same holds for many nonconstructive proofs in mathematical economics and game theory (e.g., deductions establishing the existence of equilibria using fixed-point theorems). See Lewis 1985. In summary, then, generative implies deductive, but the converse is not true.

Given the differences between agent-based modeling and both inductive and deductive social science, a distinguishing term seems appropriate. The choiceof “generative” was inspiredby Chomsky’s(1965) early usage: Syntactic theory seeks minimal rule systems that are sufficient to generate the structures of interest, grammatical constructions among them.14 The generated structures of interest here are, of course, social.

Now, at the outset, I claimed that the agent-based computational model was a scientific instrument. A fair question, then, is whether agent-based computational modeling offers a powerful new way to do empirical research.Iwill argue thatit does. Interestingly, oneof the early efforts involves the seemingly remote fields of archaeology and agent-based computation.

Empirical Agent-Based Research

The Artificial Anasazi project of Dean, Gumerman, Epstein, Axtell, Swedlund, McCarroll, and Parker aims to grow an actual 500-year spatio-temporal demographic history—the population time series and spatial settlement dynamics of the Anasazi—testing against data. The Artificial Anasazi computational model proper is a hybrid in which the physical environment is “real” (reconstructed from dendroclimatalogical and other data) and the agents are artificial. In particular, we are attemptingto modelthe Kayenta AnasaziofLong HouseValley,asmall region in northeastern Arizona, over the period 800 to 1300 AD, at which point the Anasazi mysteriously vanished from the Valley. The enigma of the Anasazi has long been a central question in Southwestern archaeology. One basic issue is whether environmental (i.e., subsistence) factors alone can account for their sudden disappearance. Or do other factors—property rights, clan relationships, conflict, disease—have to be admitted to generate the true history? In bringing agents to bear on this controversy, we have the benefitsof(a)a very accuratereconstructionof the physical environment (hydrology, aggradation, maize potential, and drought severity) on a square hectare basis for each year of the study period, and (b) an excellent reconstruction of household numbers and locations.

The logic of the exercise has been, first, to digitize the true history— we can now watch it unfold on a digitized map of Longhouse Valley. This data set (what really happened) is the target—the explanandum. The aim is to develop, in collaboration with anthropologists, micro-specifications—ethnographically plausible rules of agent behavior—that will generate the true history. The computational challenge, in other words, is to place artificial Anasazi where the true ones were in 800 AD and see if—under the postulated rules—the simulated evolution matches the true one. Is the microspecification empirically adequate, to use van Fraassen’s (1980) phrase?15 From a contemporary social science standpoint, the research also bears on the adequacy of simple “satisficing” rules—rather than elaborate optimizing ones—to account forthe observed behavior.

Acomprehensivereporton Phase1(environmental rules only)of this research is given in Dean et al. 1999. Thefull microspecification, including hypothesized agent rulesfor choosingresidences and farming plots, is elaborated there. The central result is that the purely environmental rules explored thus far account for (retrodict) important features of the Anasazi’s demography, including the observed coupling between environmental and population fluctuations, as well as important observed spatial dynamics: agglomerations and zonal occupation series. These rules also generate a precipitous decline in population around 1300. However, they do not generate the outright disappearance that occurred. One interpretation of this finding is that subsistence considerations alone do not fully explain the Anasazi’s departure, and that institutional or other cultural factors were likely involved. This work thus suggests the power and limits of a purely environmental account, a finding that advances the archaeological debate.

Simply to convey the flavor of these simulations—which unfold as animations on the computer—figure 1.1 gives a comparison.16 Each dot is an Anasazi household. The graphic shows the true situation on the right and a simulation outcome on the left for the year 1144. In both cases, agents are located at the border of the central farming area— associated with a high water table (dark shade)—and the household numbers are interestingly related.

The populationtimeseries (see Dean et al. 1999) comparing actual and simulatedfora typical runis alsorevealing. Thesimulated Anasazi curve is qualitatively encouraging, matching the turning points, including a big crash in 1300, but quantitatively inaccurate, generally overestimating population levels, and failing to generate the “extinction” event of interest.

As noted earlier, one intriguing interpretation of these results is that the Valley could have supported the Anasazi in fact, so their departure may have been the result of institutional factors not captured in the purely environmental account.

The claim is not that the current model has solved—or that the planned extensions will ultimately solve—the mystery of the Anasazi. Rather, the point is that agent-based modeling permits a new kind of empirical research (and, it might be noted, a novel kind of interdisciplinary collaboration).

This is by no means the only example of data-driven empirical research with agents. For example, Axtell (1999) gives an agent-based computational model of firm formation that generates distributions of firm sizes and growth rates close to those observed in the U.S. economy. Specifically, citing the work of Stanley et al. (1996, 806), Axtell writes that “there are three important empirical facts that an accurate theory of the firm should reproduce: (a) firm sizes must be right-skewed, approximating a power law; (b) firm growth rates must be Laplace distributed; (c) the standard deviation in log growth rates as a function of size must follow a power law with exponent −0.15 ± 0.03.” He further requires that the model be written at the level of individual human agents—that it be methodologically individualist. Aside from his own agent-based computational model, Axtell writes,“ . . . theoriesof the firm that satisfy all these requirements are unknown to us” (1999, 88).

Similarly, observed empirical size-frequency distributions for traffic jams are generated in the agent-based model of Nagel and Rasmussen (1994). Bak, Paczuski, and Shubik (1996) present an agent-based trading model that succeeds in generating the relevant statistical distribution of prices.

Axelrod(1993) develops an agent-based model of alliance formation that generates the alignment of seventeen nations in the Second World War with high fidelity. Other exercises in which agent-based models are confronted with data include Kirman andVriend 1998 and Arthur et al. 1997.

As in the case of the Anasazi work, I am not claiming that any of these models permanently resolves the empirical question it addresses. The claim, rather, is that agent-based modeling is a powerful empirical technique. In some of these cases (e.g., Axtell 1999), the agents are individual humans, and in others(Dean et al. 1999; Axelrod 1993) they are not. But, in all these cases, the empirical issue is the same: Does the hypothesized microspecification suffice to generate the observed phenomenon?—be it a stationary firm size distribution, a pattern of alliances, or a nonequilibrium price time series. The answer may be yes and, crucially, it may be no. Indeed, it is precisely the latter possibility—empirical falsifiability—that qualifies the agent-based computational model as a scientific instrument.

In addition to “hard” quantitative empirical targets, agent-based computational models may aimto generate important social phenomena qualitatively. Examples of “stylized facts” generated in such models include: right-skewed wealth distributions (Epstein and Axtell 1996), cultural differentiation (Epstein and Axtell 1996; Axelrod 1997c), multi-polarity in interstate systems (Cederman 1997), new political actors(Axelrod 1997d), epidemics (Epstein and Axtell 1996), economic classes (Axtell, Epstein, and Young 2001), and the dynamics of retirement (Axtell and Epstein 1999) to name a few. This “computational theorizing,”17 if you will, can offer basic insights of the sort exemplified in Schelling’s (1971) pioneering models of racial segregation, and may, of course, evolve into models directly comparable to data. Indeed, they may inspire the collection of data not yet in hand. (Without theory, it is not always clear what data to collect.)Turning from empirical phenomena, the generated phenomenon may be computation itself.

Connectionist Social Science

Certain social systems, such as trade networks (markets), are essentially computational architectures. They are distributed, asynchronous, and decentralized and have endogenous dynamic connectiontopologies.For example, the CD-ROM version of Epstein and Axtell 1996 presents animations of dynamic endogenous trade networks. (For other work on endogenous trade networks, seeTesfatsion 1995.) There, agents are represented as nodes, and lines joining agents represent trades. The connection pattern—computing architecture—changes as agents move about and interact economically, as shown in figure 1.2.

Whether they realize it or not, when economists say “the market arrives at equilibrium,” they are asserting that this type of dynamic “social neural net” has executed a computation—it has computed P*, an equilibrium price vector. No individual has tried to compute this, but the society of agents does so nonetheless. Similarly, convergence to social norms, convergence to strategy distributions (in n-person games), or convergence to stable cultural or even settlement patterns (as in the Anasazi case) are all social computations in this sense.

It is clear that the efficiency—indeed the very feasibility—of a social computation may depend on the way in which agents are connected. After all, information in society is not manna from heaven; it is collected and processed at the agent level and transmitted through interaction structuresthat are endogenous. How then does the endogenous connectivity—the topology—of a social network affect its performance asa distributed computational device, one that,for example, computes price equilibria, or converges to (computes) social norms, or converges to spatial settlement patterns such as cities?18 Agent-based models allow us to pursue such connectionist social science questions in new and systematic ways.

Interdisciplinary Social Science

Many important social processes are not neatly decomposable into separate subprocesses—economic, demographic, cultural, spatial—whose isolated analysis can be somehow “aggregated” to yield an adequate analysis of the process as whole. Yet this is exactly how academic social science is organized—into more or less insular departments and journals of economics, demography, anthropology, and so on. While many social scientists would agree that these divisions are artificial, they would argue that there is no “natural methodology” for studying these processes together, as they interact, though attempts have been made. Social scientists have taken highly aggregated mathematical models— of entire national economies, political systems, and so on—and have “connected” them, yielding “mega-models” that have been attacked on several grounds (see Nordhaus 1992). But attacks on specific models have had the effect of discrediting interdisciplinary inquiry itself, and this is most unfortunate. The line of inquiry remains crucially important. And agent-based modeling offers an alternative, and very natural,technique.

For example, in the agent-based model Sugarscape (Epstein and Axtell 1996), each individual agent has simple local rules governing movement, sexual reproduction, trading behavior, combat, interaction with the environment, and the transmission of cultural attributes and diseases. These rules can all be “active” at once. When an initial population of such agents is released into an artificial environment in which, and with which, they interact, the resulting artificial society unavoidably links demography, economics, cultural adaptation, genetic evolution, combat, environmental effects, and epidemiology. Because the individual is multidimensional, so is the society.

Now, obviously, not all social phenomena involve such diverse spheres of life. If one is interested in modeling short-term price dynamics in a local fish market, then human immune learning and epidemic processes may not be relevant. But if one wishes to capture long-term social dynamics of the sort discussed inWilliam McNeill’s 1976 book Plagues and Peoples, they are essential. Agent-based modelers do not insist that everything be studied all at once. The claim is that the new techniques allow us to transcend certain artificial boundaries that may limit our insight.


For example, Sugarscape agents(Epstein and Axtell 1996) engage in sexual reproduction, transmitting genes for, interalia, vision (the distance they can see in foraging for sugar). An offspring’s vision is determined by strictly Mendelian (one locus–two allele) genetics, with equal probability of inheriting the father’s or mother’s vision. One can easily plot average vision in society over time. Selection will favor agents with relatively high vision—since they’ll do better in the competition to find sugar—and, as good Darwinians, we expect to see average vision increase over time, which it does. Now, suppose we wish to study the effect of various social conventions on this biological evolution. What,for example,is the effect of inheritance—the social convention of passing on accumulated sugar wealth to offspring—on the curve of average vision? Neither traditional economics nor traditional population genetics offer particularly natural ways to study this sort of “nature-nurture” problem. But they are naturally studied in an agent-based artificial society: Just turn inheritance “off” in one run and “on” in another, and compare!19 Figure 1.3 gives a typical realization.

With inheritance, the average vision curve (gray) is lower: Inheritance “dilutes” selection. Because they inherit sugar, the offspring of wealthy agents are buffered from selection pressure. Hence, low-vision genes persistthatwouldbeselectedoutintheabsenceofthissocial convention. We do not offer this as a general law, nor are we claiming that agent-based models are the only ones permitting exploration of such topics.20 The claim is that they offer a new, and particularly natural, methodology for approaching certain interdisciplinary questions, including this one. Some of these questions can be posed in ways that subject dominant theories to stress.

Theory Stressing

One can use agent-based models to test the robustness of standard theory. Specifically, one can relax assumptions about individual—micro level—behavior and see if standard results on macro behavior collapse. For example, in neoclassical microeconomic theory, individual preferences are assumed to be fixed for the lifetime of the agent. On this assumption (and certain others), individual utility maximization leads to price equilibrium and allocative efficiency (the First Welfare Theorem). But, what if individual preferences are not fixed but vary culturally? In Epstein and Axtell 1996, we introduce this assumption into trading agents who are neoclassical in all other respects (e.g., they have Cobb-Douglas utility functions and engage only in Pareto-improving trades with neighbors). The result is far-from-equilibrium markets. The standard theory is not robust to this relaxation in a core assumption about individual behavior.Forareviewof the literature onthis central fixed preferences assumption, see Bowles 1998.

Agents, Behavioral Social Science, and the Micro-Macro Mapping

What can agent-based modeling and behavioral research do for one another? It is hard to pinpoint the dawn of experimental economics, though Simon (1996) credits Katona (1951) with the fundamental studies of expectation formation. In any event, there has been a resurgence of important laboratory and other experimental work on individual decision making over the last two decades. See, for example, Camerer 1997, Rabin 1998, Camerer and Thaler 1995, Tversky and Kahneman 1986, and Kagel and Roth 1995. This body of laboratory social science, if I may call it that, is giving us an ever-clearer picture of how homo sapiens—as against homo economicus—actually makes decisions. However, a crucial lesson of Schelling’s segregation model, and of many subsequent Cellular Automaton models, such as “Life”—not to mention agent-based models themselves—is that even perfect knowledge of individual decision rules does not always allow us to predict macroscopic structure. We get macro-surprises despite complete micro-knowledge. Agent-based models allow us to study the micro-to-macro mapping. It is obviously essential to begin with solid foundations regarding individual behavior,and behavioral researchisclosinginonthese. However,wewill still need techniques for “projecting up” to the macro level from there (particularly for spatially-distributed systems of heterogeneous individuals). Agent modeling offers behavioral social science a powerful way to do that.

Agent-based models may also furnish laboratory research with counterintuitive hypotheses regarding individual behavior. Some, apparently bizarre, system of individual agent rules may generate macrostructures that mimic the observed ones. Is it possible that those are, in fact, the operative micro-rules? It might be fruitful to design laboratory experiments to test hypotheses arising from the unexpected generative sufficiency of certain rules.

What does behavioral research offer agent-based modeling? Earlier, we noted that different agent-based models might have equal generative (explanatory) power and that, in such cases, further work would be necessary to adjudicate between them. But if two models are doing equally well in generating the macrostructure, preference should go to the one that is best at the micro level. So, if we took the two microspecifications as competing hypotheses about individual behavior, then—apropos of the preceding remark—behavioral experiments might be designed to identify the better hypothesis (microspecification) and, in turn, the better agent model. These, then, arefurther ways in which agent-based computational modeling can contribute to empirical social science research.


As noted earlier, to adopt agent-based modeling does not compel one to adopt methodological individualism. However, extreme methodological individualism is certainly possible (indeed common) in agent-based models. And individual-based models may have the important effect of decoupling individual rationality from macroscopic equilibrium. For example, in the individual-based retirement model of Axtell and Epstein (1999), macroscopic equilibrium is attained—through a process of imitation in social networks—even though the vast preponderance of individuals are not rational. Hence—as in much evolutionary modeling— micro rationalityis nota necessary conditionforthe attainmentof macro equilibrium.21 Now, we also have agent-models in which macroscopic equilibrium is not attained despite orthodox utility maximization at the individual level. The non-equilibrium economy under evolving preferences (Epstein and Axtell 1996) noted earlier is an example. Hence, micro rationality is not a sufficient condition for macro equilibrium. But if individual rationality is thus neither necessary nor sufficient for macro equilibrium, the two are logically independent—or decoupled, if you will.

Now, the fraction of agents in an imitative system (such as the retirement model) who are rational will definitely affect the rate at which any selected equilibrium sets in. But the asymptotic equilibrium behavior per se doesnotdependonthedialofrationality,despitemuch behavioral research on this latter topic. Perhaps the main issue is not how much rationality there is (at the micro level), but how little is enough to generate the macro equilibrium.

In passing, it is worth noting that this is of course a huge issue for policy, where “fad creation” may be far more effective than real education. Often, the aim is not to equip target populations with the data and analytical tools needed to make rational choices; rather, one displays exemplars and then presses for mindless imitation. “Just say no to drugs” not because it’srational—ina calculusof expected lifetime earnings—but because a famous athlete says “no” and it’s a norm to imitate him. The manipulation of uncritical imitative impulses may be more effective in getting to a desired macro equilibrium than policies based on individual rationality.Thesocial problem,of course,isthat populationsof uncritical imitators are also easy fodder for lynch mobs, witch hunts, Nazi parties, and so forth. Agent-based modeling is certainly not the only way to study social contagion (see, for example, Kuran 1989), but it is a particularly powerful way when the phenomenon is spatial and the population in question is heterogeneous.

Relatedly, agent-based approaches may decouple social science from decision science. In the main, individuals do not decide—they do not choose—in any sensible meaning of that term, to be ethnic Serbs, to be native Spanish speakers, or to consider monkey brain a delicacy. Game theory may do an interesting job explaining the decision of one ethnic group to attack another at a certain place or time, but it doesn’t explain how the ethnic group arises in the first place or how the ethnic divisions are transmitted across the generations. Similarly for economics, what makes monkey brain a delicacy in one society and not in another? Cultural (including preference) patterns, and their nonlinear tippings, are topics of study in their own right with agents.22 See Axelrod 1997band Epstein and Axtell 1996.

Analytical-Computational Approach to Non-Equilibrium Social Systems

For many social systems, it is possible to prove deep theorems about asymptotic equilibria. However, the time required for the system to attain (or closely approximate) such equilibria can be astronomical. The transient, out-of-equilibrium dynamics of the system are then of fundamental interest. A powerful approach is to combine analytical proofs regarding asymptotic equilibria with agent-based computational analyses of long-lived transient behaviors, the meta-stability of certain attractors, and broken ergodicity in social systems.

One example of this hybrid analytical-computational approach is Axtell, Epstein, andYoung 2001.We develop an agent-based model to study the emergence and stability of equity norms in society. (In that article, we explicitly define the term “emergent” to mean simply “arising from decentralized bilateral agent-interactions.”) Specifically, agents with finite memory play Best Reply to Recent Sample Evidence (Young 1995, 1998) in a three-strategy Nash Demand Game, and condition on an arbitrary “tag” (e.g., a color) that initially has no social or economic significance—it is simply a distinguishing mark. Expectations are generated endogenously through bilateral interactions. And, over time, these tags acquire socially organizing salience. In particular, tag-based classes arise.(The phenomenonis akintothe evolutionof meaning discussedin Skyrms 1998.)

Now, introducing noise, it is possible to cast the entire model as a Markov process and to prove rigorously that it has a unique stationary strategy distribution. When the noise level is positive and sufficiently small, the following asymptotic result can be proved: The state with the highest long-run probability is the equity norm, both between and within groups.

Salutary as this asymptotic result may appear, the transition from inequitable states to these equitable ones can be subject to tremendous inertia. Agent-based models allow us to systematically study long-lived transient behaviors.We know that, beginningin an inequitable regime, the system will ultimately “tip” into the equity norm. But how does the waiting time to this transition depend on the number of agents and on memory length? In this case, the waiting time scales exponentially in memory length, m, and exponentially in N, the number of agents. Overall, then, the waiting time is immense for m= 10 and merely N= 100, for example.

Speaking rigorously,the equity normis stochastically stable(seeYoung 1998). The agent-based computational model reveals, however, that— depending on the number of agents and their memory lengths—the waiting time to transit from an inequitable regime to the equitable one may be astronomically long.

This combination of formal (asymptotic) and agent-based (non-equilibrium) analysis seems to offer insights unavailable from either approach alone, and to represent a useful hybrid form of analytical-computational study. For sophisticated work relating individual-based models to analytical ones in biology, see Flierl et al. 1999.

Foundational Issues

We noted earlier that markets can be seen as massively parallel spatially distributed computational devices with agents as processing nodes. To say that “the market clears” is to say that this device has completed a computation. Similarly, convergence to social norms, convergence to strategy distributions (in n-person games), or convergence to stable cultural or settlement patterns, are all social computations in this sense. Minsky’s (1985) famous phrase was “the Society of Mind.” What I’m interested in here is “the Society as Mind,” society as a computational device. (On that strain of functionalism which would be involved in literally asserting that a society could be a mind, see Sober 1996.)

Now, once we say “computation” we think of Turing machines (or, equivalently, of partial recursive functions). In the context of n-person games, for example, the isomorphism with societies is direct: Initial strategies aretallies onaTuring machine’s input tape; agent interactions function to update the strategies (tallies) and thus represent the machine’s state transition function; an equilibrium is a halting state of the machine; the equilibrium strategy distribution is given by the tape contents in the halting state; and initial strategy distributions that run to equilibrium are languages accepted by the machine. The isomorphism is clear. Now, we know what makes for an intractable, or “hard,” computational problem. So, given our isomorphism, is there a computational answer to the question, “What’s a hard social problem?”

AComputational Characterization of Hard Social Problems

In the model of tag-based classes discussed earlier (Axtell, Epstein, and Young 2001), we prove rigorously that, asymptotically, the equity norm will set in. However, beginning from any other (meta-stable) equilibrium, the time to transit into the equitable state scales exponentially in the number of agents and exponentially in the agents’ memory length. If we adopt the definition that social states are hard to attain if they are not effectively computable by agent society in polynomial time, then equity is hard. (The point applies to this particular setup; I am emphatically not claiming that there is anything immutable about social inequity.) In a numberof models, the analogous point appliesto economic equilibria: There are nonconstructive proofs of their existence but computational arguments that their attainment requires time that scales exponentially in, for instance, the dimension of the commodity space.23 On our tentative definition, then, computation of (attainment of) economic equilibria would qualify as another hard social problem.

So far we have been concerned with the question, “Does an initial social state run to equilibrium?” or, equivalently, “Does the machine halt given input tape x?” Now, like satisfiability, or truth-table validity in sentential logic, these problems are in principle decidable (that is, the equilibria are effectively computable), but not on time scales of interest to humans. (Here, with Simon [1978], we use the term “time” to denote “the number of elementary computation steps that must be executed to solve the problem.”)

Gödelian Limits

But there are social science problems that are undecidable in principle, nowin thesenseof Gödel or the Halting Problem. Rabin(1957) showed that “there are actual win-lose games which are strictly determined for which there is no effectively computable winning strategy.” He continues, “Intuitively, our result means that there are games in which the player who in theory can always win, cannot do so in practice because it is impossible to supply him with effective instructions regarding how he should play in order to win.” Another nice example, based on the unsolvabilityof Hilbert’sTenth Problem,is givenby Prasad (1997):

For n-player games with polynomial utility functions and natural number strategy sets the problem of finding an equilibrium is not computable. There does not exist an algorithm which will decide, for any such game, whether it has an equilibrium or not . . . When the classof gamesis specifiedbya finite set of players, whose choice sets are natural numbers, and payoffs are given by polynomial functions, the problem of devising a procedure which computes Nash equilibria is unsolvable.

Other resultsof comparablestrength have been obtainedby Lewis(1985, 1992a, and 1992b).24

Implications for Rational Choice Theory

Here lies the deepest conceivable critique of rational choice theory. There are strategic settings in which the individually optimizing behavior is uncomputable in principle. A second powerful critique is that, while possible in principle, optimization is computationally intractable. As Duncan Foley summarizes, “The theory of computability and computational complexity suggest that there are two inherent limitations to the rational choice paradigm. One limitation stems from the possibility that the agent’s problem is in fact undecidable, so that no computational procedure exists which for all inputs will give her the needed answer in finite time.Asecond limitationis posedby computational complexityin that even if her problem is decidable, the computational cost of solving it may in many situations be so large as to overwhelm any possible gains from the optimal choice of action” (see Albin 1998, 46). For a fundamental statement, see Simon 1978.

These possibilities are disturbing to many economists. They implicitly believe that if the individual is insufficiently rational it must follow that decentralized behavior is doomed to produce suboptimality at the aggregate level. The invisible hand requires rational fingers, if you will. There are doubtless cases in which this holds. But it is not so in all cases. As noted earlier, in the retirement model of Axtell and Epstein (1999), as well as in much evolutionary modeling, an ensemble of locally interacting agents—none of whom are canonically rational—can nonetheless attain efficiency in the aggregate. Even here, of course, issues of exponential waiting time arise (as in the classes model above). But it is important to sort the issues out.

The agent-based approach forces on us the interpretation of society as a computational device, and this immediately raises foundational specters of computational intractability and undecidability. Much of the economic complexity literature concerns the uncomputability of optimal strategies by individual rational agents, surely an important issue. However, our central concern is with the effective computability (attainment) of equilibria by societies of boundedly rational agents. In that case, it is irrelevant that equilibrium can be computed by an economist external to the system using the Scarf, or other such, algorithm. The entire issue is whether it can be attained—generated— through decentralized local interactions of heterogeneous boundedly rational actors. And the agent-based computational model is a powerful tool in exploring that central issue. In some settings, it may be the only tool.

Equations versus Agent-Based Models

Three questions arise frequently and deserve treatment: Given an agent-based model, are there equivalent equations? Can one “understand” one’s computational model without such equations?If one has equations for the macroscopic regularities, why does one need the “bottom-up” agent model?

Regarding the first question—are there equivalent equations for every computational model—the answer is immediate and unequivocal: absolutely. On the Church-Turing Thesis, every computation (and hence every agent-based model) canbe implementedbyaTuring machine. For every Turing machine there is a unique corresponding and equivalent Partial Recursive Function (see Rogers 1967). Hence, in principle, for any computation there exist equivalent equations (involving recursive functions). Alternatively, any computer model uses some finite set of memory locations, which are updated as the program executes. One can think of each location as a variable in a discrete dynamical system. In principle, there is some—perhaps very high dimensional—set of equations describing those discrete dynamics. Now, could a human write the equations out? Solve them or even find their equilibria(if such exist)? The answer is not clear. If the equations are meant to represent large populations of discrete heterogeneous agents coevolving on a separate space, with which they interact, it is not obvious how to formulate the equations, or how to solve them if formulated. And, for certain classes of problems(e.g.,the PSPACE Complete problems), it can be proved rigorously that simulation is—in a definite sense—the best one can do in principle (see Buss, Papadimitriou, andTsitsiklis 1991). But that does not mean—turning to the second question—we have no idea what’s going on in the model.

To be sure, a theorem is better than no theorem. And many complex social phenomena may ultimately yield to analytical methods of the sort being pioneered by Young (1998), Durlauf (1997b), and others. But an experimental attitude is also appropriate. Consider biology. No one would fault a “theoremless” laboratory biologist for claiming to understand population dynamics in beetles when he reports a regularity observed over a large number of experiments. But when agent-based modelers show such results—indeed, far more robust ones—there’s a demand for equations and proofs. These would be valuable, and we should endeavor to produce them. Meanwhile, one can do perfectly legitimate “laboratory” science with computers, sweeping the parameter space of one’s model, and conducting extensive sensitivity analysis, and claiming substantial understanding of the relationships between model inputs and model outputs, just as in any other empirical science for which general laws are not yet in hand.25

Thethird question involves confusion between explanation and description, and might best be addressed through an example. In Epstein and Axtell 1996, spatially distributed local agent interactions generate the oscillatory aggregate population time series shown in figure 1.4.

The question then arises: Could you not get that same curve from some low-dimensional differential equation, and if so, why do you need the agent model? Let us imagine that we can formulate and analytically solve such an equation, and that the population trajectory is exactly P(t) = A+ BSin(Ct) for constants A, B, and C. Now, what is the explanatory significance of that descriptively accurate result? It depends on one’s criteria for explanation. If we are generativists, the question is: How could the spatially decentralized interactions of heterogeneous autonomous agents generatethat macroscopicregularity? If that is one’s question, then the mere formula P(t) = A+ BSin(Ct) is devoid of explanatory power despite its descriptive accuracy. The choice of agents versus equations always hinges on the objectives of the analysis. Given some perfectly legitimate objectives, differential equations are the tool of choice; given others, they’re not. If we are explicit as to our objectives, or explanatory criteria, no confusion need arise. And it may be that hybrid models of a second sort are obtainable in which the macrodynamics are well described by an explicit low-dimensional mathematical model, but are also generated from the bottom up in a model population of heterogeneous autonomous agents. That would be a powerful combination. In addition to important opportunities, the field of agent-based modeling, like any young discipline, faces a number of challenges.


First, the field lacks standards for model comparison and replication of results; see Axtell et al. 1996. Implicit in this is the need for standards in reportage of assumptions and certain procedures. Subtle differences can have momentous consequences. For example, how, exactly, are agents being updated? The Huberman and Glance (1993) critique of NowakandMay(1992)is strikingproofthat asynchronous updatingof agents produces radically different results from synchronous updating. Huberman and Glance show that Nowak and May’s main result—the persistence of cooperation in a spatial Prisoner’s Dilemma game— depends crucially on synchronous updating. When, ceteris paribus, Huberman and Glance introduce asynchronous updating into the Nowak and May model, the result is convergence to pure defection. (For a spatial Prisoner’s Dilemma model with asynchronous updating in which cooperation can persist, see Epstein 1998.) The same sorts of issues arise in randomizing the agent call order, where various methods—with different effects on output—are possible.

It is also fair to say that solution concepts are weak. Certainly, hitting the “Go” button and watching the screen does not qualify as solving anything—any more than an eveningat the casino solves the Gambler’s Ruin Problem from Markov Theory. An individual model run offers a sample path of a (typically) stochastic process, but that is not a general solution—a specific element of some well definedfunction space (e.g., a Hilbert or Sobolev space). As noted earlier, it is often possible to sweep the parameter space of one’s model quite systematically and thereby obtain a statistical portrait of the relationship between inputs and outputs, asin Axelrod 1997c or Epstein 1998. Butitisfairtosay that this practice has not been institutionalized.

A deeper issue is that sweeping a model’s numerical parameter space is easier than exploring the space of possible agent behavioral rules (e.g., “If a neighboring agent is bigger than you, run away” or “Always share food with kin agents”). For artificial societies of any complexity (e.g., Sugarscape), we have no efficient method of searching the space of possible individual rules for those that exhibit generative power. One can imagine using evolutionary approaches to this. First, one would define a metric such that, given a microspecification, the distance from model outputs (generated macrostructures) to targets (observed macrostructures) could be computed. The better the match (the smaller this distance) the “fitter” is the microspecification. Second, one would encode the space of candidate micro specifications and turn, say, a Genetic Algorithm (GA) (see Holland 1992; Mitchell 1998) loose on it. The GA might turn up counterintuitive boundedly rational rules that are highly “fit” in this sense of generating macrostructures “close” to the targets. (These then become hypotheses for behavioral research, as discussed earlier.)

This strikes me as a far more useful application of GAs than the usual one: finding hyper-rational individual strategies, which we now have strong experimental evidence are not being employed by humans. The problem is how to encode the vast space of possible individual rules (not to mention the raw computational challenge of searching it once encoded).Insomerestricted cases,this has been done successfully (Axelrod 1987; Crutchfield and Mitchell 1995), but for high dimensional agents engaged in myriad social interactions—economic, cultural, demographic—it is far from clear how to proceed.

One of the central concepts in dynamics is sensitivity. Sensitivity involves the effect on output (generated macrostructure) of small changes in input (microspecification). To assess sensitivity in agent models, we have to do more than encode the space of rules—we have to metrize it. To clarify the issue, consider the following agent rules(methods of agent-objects):

Rule a = Never attack neighbors.
Ruleb = Attacka neighborifhe’s green.
Rule c = Attack a neighbor if he’s smaller than you.

Which rule—b or c—represents a “smaller departure from” Rule a? Obviously, the question is ill-posed. And yet we speak of “small changes in the rules” of agent-based models. Some areas (e.g., Cellular Automata) admit binary encodings of rule space for which certain metrics—taxicab or Hamming distance—are natural. But for artificial societies generally, no such simple avenues present themselves. What then constitutesa small rule change? Without some metric, we really cannot develop the analogue, for agent-based models, of structural stability—or equivalently, of bifurcation theory—in dynamical systems.

Some challenges are sociological. Generating collective behavior that to the naked eye “looks like flocking” can be extremely valuable, but it is aradically different enterprise from generating, say,aspecific distribution of wealth with parameters close to those observed in society. Crude qualitative caricature is a perfectly respectable goal. But if that is one’s goal, the fact must be stated explicitly—perhaps using the terminology proposed in Axtell and Epstein 1994. This will avert needless resistance from other fields where “normal science” proceeds under established empirical standards patently not met by cartoon “boid” flocks, however stimulating and pedagogically valuable these may be. On the pedagogical value of agent-based simulation generally, see Resnick 1994.

Anumber of other challenges include building community and sharing results and are covered in Axelrod 1997a. In addition to foundational, procedural, and other scientific challenges, the field of “complexity” and agent-based modelingfaces terminological ones.In particular,the term “emergence” figures very prominently in this literature. It warrants an audit.


I have always been uncomfortable with the vagueness and occasional mysticism surrounding this word and, accordingly, tried to define it quite narrowly in Epstein and Axtell 1996. There, we defined “emergent phenomena” to be simply “stable macroscopic patterns arising from local interaction of agents.”26 Many researchers define the term in the same straightforward way (e.g., Axelrod 1997a). Since our work’s publication, I have researched this term more deeply and find myself questioning its adoption altogether.

“Emergence”hasa history,anditisan extremely spotty one, beginning with classical British emergentism in the 1920s and the works of Samuel Alexander (Space, Time, and Deity, 1920), C. D. Broad (The Mind and Its Place in Nature, 1925), and C. Lloyd Morgan (Emergent Evolution, 1923). The complexity community should be alerted to this history. There is an unmistakably anti-scientific—even deistic—flavor to this movement, which claimed absolute unexplainability for emergent phenomena. In the view of these authors, emergent phenomena are unexplainable in principle. “The existence of emergent qualities . . . admits no explanation,” wrote Alexander (1920).27 As philosopher Terence Horgan recounts, emergent phenomena were to be “accepted (in Samuel Alexander’s striking phrase) ‘with natural piety.’”

Striking indeed, this sort of language, and classical emergentism’s avowedly vitalist cast (see Morgan 1923) stimulated a vigorous—and to my mind, annihilative—attack by philosophers of science. In particular, Hempeland Oppenheim(1948) wrote, “This versionof emergence . . . is objectionable not only because it involves and perpetuates certain logical confusions but also because not unlike the ideas of neovitalism, it encourages an attitude of resignation which is stifling to scientific research. No doubt it is this characteristic, together with its theoretical sterility, which accounts for the rejection, by the majority of contemporary scientists, of the classical absolutist doctrine of emergence.”

Classical absolute emergentism is encapsulated nicely in the following formalization of Broad’s(1925, 61):

Put in abstract terms the emergent theory asserts that there are certain wholes, composed (say) of constituents A, B, and C in a relation R to each other . . . and that the characteristic propertiesof the whole R(A,B,C) cannot, even in theory, be deduced from the most complete knowledge of the propertiesofA,B,andCin isolationorinother wholeswhicharenot in the form R(A,B,C). (Emphasis in original)

Before explicating the logical confusion noted by Hempel and Oppenheim, we can fruitfully apply a bit of logic ourselves. Notice that we have actually accumulated a number of first-order propositions. For predicates, let C stand for classically emergent, Dfor deducible, E for explained, and Gfor generated (in a computational model). Then, if x is a system property, we have:

(1)    (∀x)(Cx ⊃¬Dx)     Broad (emergent implies not deducible)28
(2)    (∀x)(Cx ⊃¬Ex)     Alexander (emergent implies not explainable)
(3)    (∀x)(¬Gx ⊃¬Ex)     Generativist Motto (not generated implies not explained)
(4)    (∀x)(Gx Dx)     Theorem (generated implies deduced)

Although a number of derivations are possible,29 the essential point involves (1) and (4). By the earlier Theorem (4), if x is generable, then it is deducible. But, by Broad (1), if x is emergent, it is not deducible. But it then follows that if x is generable, then it cannot be emergent!30 In particular, if x is generated in an agent-based model, it cannot be classically emergent. Agent-based modeling and classical emergentism are incompatible. Further incompatibilities between agent-based modeling and classical emergentism will be taken up below.

Logical Confusion

Now, the logical confusion noted earlier is set forth clearly in Hempel and Oppenheim 1948, is discussed at length in Nagel 1961, and is recounted more recently by Hendriks-Jansen 1996. To summarize, like Broad, emergentists typically assert things like, “One cannot deduce higher properties from lower ones; macro properties from micro ones; the properties of the whole from the parts’ properties.” But, we do not deduce properties. We deduce propositions in formal languages from other propositions in those languages.31 This is not hair-splitting: If the macro theory contains terms (predicates, variable names) that are not terms of the micro theory, then of course it is impossible to deduce macro claims involving those terms from propositions of the micro theory. It is logically impossible. So the “higher emergent” property of water, “translucence,” is trivially not deducible from the micro theory of oxygen (O) and hydrogen (H) since “translucent” is not a term of the micro theory. Many so called “emergent properties” of “wholes” are not deducible from “parts”for this purely logical reason. So emergence, as nondeducibility, is always relative to some theory (some set of well-formed formulae and inference rules); it is not absolute as the classicals would have it.

Arelative versionof emergencedueto Hempeland Oppenheim(1948) is formalized in Stephan 1992 as follows. Consider a system with constituents C1, . . . ,Cn in relation O to one another (analogous to Broad’s A, B, C, and R). “This combination is termed a microstructure [C1, . . . Cn;O]. And let T be a theory. Then, a system property P is emergent, relative to this microstructure and theoryT, if:

(a). There is a law LP which holds: for all x, when x has microstructure [C1 . . . . Cn;O] thenx has propertyP, and
(b). By means of theoryT, LP cannot be deduced from laws governing the Cl . . . . Cnin isolationorin other microstructuresthanthe given.”

Stephan continues, “By this formulation the original absolute claim has been changed into a merely relative one which just states that at a certain time according to the available scientific theories we are not able to deduce the so-called emergent laws” (1992, 39).32 But now, as Hempel and Oppenheim write, “If the assertion that life and mind have an emergent status is interpreted in this sense, then its import can be summarized approximately by the statement that no explanation, in terms of microstructure theories, is available at present for large classes of phenomena studied in biology and psychology” (emphases added). This quite unglamorous point,they continue, would “appear to represent the rational core of the doctrine of emergence.” Not only does this relative formulation strip the term of all higher Gestalt harmonics, but it suggests that,for any given phenomenon, emergent status itself maybe fleeting.

As scientific theories progress, in other words, that which was unexplainableand “emergent” ceasestobe.The chemical bond—afavoriteof the British emergentists—is an excellent example. Here,Terrence Horgan (1993) is worth quoting at length:

When Broad wrote, “Nothing that we know about Oxygen by itself or in its combination with anything but Hydrogen would give us the least reason to suppose that it would combine with Hydrogen at all. Nothing that we know about Hydrogen by itself or in its combinations with anything but Oxygen would give us the least reason to expect that it would combine with Oxygen at all” (1925, pp. 62–63), his claim was true. Classical physics could not explain chemical bonding. But the claim didn’t stay true for long: by the end of the decade quantum mechanics had come into being, and quantum-mechanical explanations of chemical bonding were in sight.

The chemical bond no longer seemed mysterious and “emergent.” Another example was biology,for the classical emergentistsarichsource of higher “emergent novelties,” putatively unexplainable in physical terms. Horgan continues,

Within another two decades, James Watson and Francis Crick, drawing upon the work of Linus Pauling and others on chemical bonding, explained the information-coding and self-replicating properties of the DNA molecule, thereby ushering in physical explanations of biological phenomena in general.

As he writes, “These kinds of advances in science itself, rather than any internal conceptual difficulties, were what led to the downfall of British emergentism, as McLaughlin (1992) persuasively argues.” Or, as Herbert Simon (1996) writes, “Applied to living systems the strong claim [quoting the “holist” philosopher J. C. Smuts] that ‘the putting together of their parts will not produce them or account for their characters and behaviors’ implies a vitalism that is wholly antithetical to modern molecular biology.”

In its strong classical usage, the term “emergent” simply “baptizes our ignorance,” to use Nagel’s phrase(1961, 371). And, when de-mystified, it can mean nothing more than “not presently explained.” But, this is profoundly different from “not explainable in principle,” as Alexander and his emergentist colleagues would have it, which is stifling, not to mention baseless empirically. As Hempel and Oppenheim wrote,

Emergence is not an ontological trait inherent in some phenomena; rather it is indicative of the scope of our knowledge at a given time; thus it has no absolute, but a relative character; and what is emergent with respect to the theories available today may lose its emergent status tomorrow. (1948, 263)

Good Questions

Now, all the questions posed by agent-based modelers and complexity scientists in this connection are fine: How do individuals combine to form firms, or cities, or institutions, or ant colonies, or computing devices? These are all excellent questions. The point is that they are posable— indeed most productively posed—without the imprecise and possibly self-mystifying terminology of “emergence,” or “supervenience,” as Morgan called it. Obviously, “wholes” may have attributes or capabilities that their constituent parts cannot have (e.g., “whole” conscious people can have happy memories of childhood while, presumably, individual neurons cannot). Equally obvious,the parts havetobe hookedup right— or interact in specific, and perhaps complicated, ways—for the wholeto exhibit those attributes.33 We at present may be able to explain why these specific relationships among parts eventuate in the stated attributes of wholes, and we may not. But, unlike classical emergentists, we do not preclude such explanation in principle.

Indeed,by attemptingto generatethese very phenomenaon computers or in mathematical models, we are denying that they are unexplainable or undeducible in principle—we’re trying to explain them precisely by figuring out microrules that will generate them. In short, we agent-based modelers and complexity researchers actually part company with those, like Alexander and company, whoseterminology we have, perhaps unwittingly, adopted. Lax definitions can compound the problem.

Operational Definitions

Typical of classical emergentism would be the claim: No description of the individual bee can ever explain the emergent phenomenon of the hive. How would one know that? Is this a falsifiable empirical claim, or something that seems true because of a lax definition of terms? Perhaps the latter. The mischievous piece of the formulation is the phrase “description of the individual bee.” What is that? Does “the bee’s” description not include its rules for interacting with other bees? Certainly, it makes little sense to speak of a Joshua Epstein devoid of all relationships with family, friends, colleagues, and so forth. “Man is a social animal,” quoth Aristotle. My “rules of social interaction” are, in part, what make me me. And, likewise, the bee’s interaction rules are what make it a bee—and not a lump. When (as a designer of agent objects) you get these rules right—when you get “the individual bee” right—you get the hive, too. Indeed, from an operationist (Hempel 1956) viewpoint, “the bee” might be defined as that x which, when put together with other x’s, makes the hive (the “emergent entity”). Unless the theoretical (model) bees generate the hive when you put a bunch of them together, you haven’t described “the bee” adequately. Thus, contrary to the opening emergentist claim, it is precisely the adequate description of “the individual bee” that explains the hive. An admirable modeling effort along precisely such lines is Theraulaz, Bonabeau, and Deneubourg 1998.

Agent-Based Modeling Is Reductionist

Classical emergentism holds that the parts (the microspecification) cannot explain the whole (the macrostructure), while to the agent-based modeler, it is precisely the generative sufficiency of the parts (the microspecification) that constitutes the whole’s explanation! In this particular sense, agent-based modeling is reductionist.34 Classical emergentism seeks to preserve a “mystery gap” between micro and macro; agent-based modeling seeks to demystify this alleged gap by identifying microspecifications that are sufficient to generate—robustly and replicably—the macro (whole). Perhaps the following thoughts of C. S. Peirce (1879) are apposite:

One singular deception . . . which often occurs,isto mistakethe sensation produced by our own unclearness of thought for a character of the object we are thinking. Instead of perceiving that the obscurity is purely subjective, we fancy that we contemplate a quality of the object which is essentially mysterious; and if our conception be afterward presented to us in a clear form we do not recognize it as the same, owing to the absence of the feeling of unintelligibility.

Explanation and Prediction

Afinalpointisthat classical emergentism trafficsona crucial(andtothis day quite common) confusion: between explanation and prediction. It may well be that certain phenomena are unpredictable in principle(e.g., stochastic). But that does not mean—as classical emergentists would have it—that they are unexplainablein principle. Platetectonics explains earthquakes but does not predict their occurrence; electrostatics explains lightning but does not predict where it will hit; evolutionary theory explains species diversity but does not predict observed phenotypes. In short, one may grant unpredictability without embracing “emergence,” as absolute unexplainability, à la Alexander and colleagues.35 And, of course, it may be that in some cases prediction is a perfectly reasonable goal. (For further distinctions between prediction and explanation, see Scheffler 1960; Suppes 1985; and Newton-Smith 1981.)

In its strong classical usages—connoting absolute nondeducibility and absolute unexplainability—“emergentism” is logically confused and antiscientific. In weak level-headed usages—like “arising from local agent interactions”—a special term hardly seems necessary. For other attempts to grapple with the term “emergent,” see Cariani 1992, Baas 1994, Gilbert 1995, and Darley 1996. At the very least, practitioners—and I include myself—should define this term carefully when they use it and distinguish their, perhaps quite sensible, meaning from others with which the term is strongly associated historically. To anyone literate in the philosophy of science, “emergence” has a history, and it is one with which many scientists may—indeed should—wish to part company. Doubtless, my own usage has been far too lax, so this admonition is directed as much at myself as at colleagues.

Recapitulation and Conclusion

I am not a soldier in an agent-based methodological crusade. For some explanatory purposes, low dimensional differential equations are perfect. For others, aggregate regression is appropriate. Game theory offers deep insight in numerous contexts and so forth. But agent-based modeling is clearly a powerful tool in the analysis of spatially distributed systems of heterogeneous autonomous actors with bounded information and computing capacity. It is the main scientific instrument in a generative approach to social science, and a powerful tool in empirical research. It is well suited to the study of connectionist phenomena in social science. It offers a natural environment for the study of certain interdisciplinary questions. It allows us to test the sensitivity of theories, such as neoclassical microeconomics, to relaxations in core assumptions (e.g., the assumption of fixed preferences). It allows us to trace how individual (micro) rules generate macroscopic regularities. In turn, we can employ laboratory behavioral research to select among competing multiagent models having equal generative power. The agent-based approach may decouple individual rationality from macroscopic equilibrium and separate decision science from social science more generally. It invites a synthesis of analytical and computational perspectives that is particularly relevant to the study of non-equilibrium systems. Agent-based models have significant pedagogical value. Finally, the computational interpretation of social dynamics raises foundational issues in social science— some related to intractability, and some to undecidability proper. Despite a number of significant challenges, agent-based computational modeling can make major contributions to the social sciences.

Return to Book Description

File created: 8/7/2007

Questions and comments to:
Princeton University Press

New Book E-mails
New In Print
PUP Blog
Princeton APPS
Sample Chapters
For Reviewers
Class Use
Recent Awards
Princeton Shorts
Freshman Reading
PUP Europe
About Us
Contact Us
PUP Home

Bookmark and Share