From Factors to Actors: Computational Sociology and Agent-Based Modeling

Author(s): Michael W. Macy and Robert Willer

Reviewed work(s):
Source: Annual Review of Sociology, Vol. 28 (2002), pp. 143-166
Published by: Annual Reviews
Stable URL: http://www.jstor.org/stable/3069238 .
Accessed: 10/06/2012 03:40

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact [email protected].

Annual Reviews is collaborating with JSTOR to digitize, preserve and extend access to Annual Review of
Sociology.

http://www.jstor.org
 Annu. Rev. Sociol. 2002. 28:143-66
doi: 10.1146/annurev.soc.28.110601.141117
Copyright? 2002 by AnnualReviews. All rightsreserved

TOACTORS:Computational
FROMFACTORS
Sociology and Agent-BasedModeling
MichaelW. Macyand RobertWiller
Departmentof Sociology, Cornell University,Ithaca, New York84153;
e-mail: [email protected],[email protected]

Key Words simulation,complexity,emergence,self-organization, cellular

automata,geneticalgorithm
* Abstract Sociologistsoften model social processesas interactionsamongvari-
ables.Wereviewan alternativeapproachthatmodelssociallife as interactionsamong
adaptiveagentswho influenceone anotherin responseto the influencethey receive.
Theseagent-basedmodels(ABMs)showhow simpleandpredictablelocalinteractions
cangeneratefamiliarbutenigmaticglobalpatterns,suchas thediffusionof information,
emergenceof norms,coordinationof conventions,orparticipation in collectiveaction.
Emergentsocial patternscan also appearunexpectedlyandthenjust as dramatically
transformor disappear,as happensin revolutions,marketcrashes,fads, and feeding
frenzies.ABMs providetheoreticalleveragewherethe globalpatternsof interestare
morethanthe aggregationof individualattributes,butat the sametime,the emergent
patterncannotbe understoodwithouta bottomup dynamicalmodelof the microfoun-
dationsat the relationallevel. Webeginwith a briefhistoricalsketchof the shiftfrom
"factors"to "actors"in computationalsociologythatshowshowagent-basedmodeling
differsfundamentally fromearliersociologicaluses of computersimulation.Wethen
reviewrecentcontributions focusedon theemergenceof socialstructureandsocialor-
deroutof localinteraction.Althoughsociologyhaslaggedbehindothersocialsciences
in appreciatingthisnew methodology,a distinctivesociologicalcontribution is evident
in the paperswe review.First,theoreticalinterestfocuseson dynamicsocialnetworks
thatshapeand are shapedby agentinteraction.Second,ABMs are used to perform
virtualexperimentsthattestmacrosociological theoriesby manipulating structuralfac-
tors like networktopology,social stratification,or spatialmobility.We concludeour
reviewwith a seriesof recommendations for realizingthe rich sociologicalpotential
of this approach.

INTRODUCTION: AGENT-BASED MODELS AND

SELF-ORGANIZING GROUP PROCESSES

Consider a flock of geese flying in tight formation. Collectively they form the
image of a giant delta-shapedbirdthatmoves as purposivelyas if it were a single
organism.Yet the flock has no "groupmind"nor is there a "leaderbird"chore-
ographingthe formation(Resnick 1997). Rather,each birdreactsto the movement
0360-0572/02/0811-0143$14.00 143
144 MACY ? WILLER

of its immediateneighborswho in turnreactto it. The resultis the gracefuldance-

like movement of the flock whose hypnotic rhythmis clearly patternedyet also
highly nonlinear.
If we tried to model the elegance of the flock at the global1 level, the effort
would be misleadingbecausethe flock is not governedby a system-wideprogram.
Moreover,thetaskwouldbe immenselydifficultbecauseof theextremecomplexity
of a nonlinearsystem. Yet the task turnsout to be remarkablyeasy if instead we
model the flock as the aggregationof local interactions.This was demonstrated
by Craig Reynolds (1987) when he modeled the movement of a population of
artificial"boids"based on three simple rules:
* Separation:Don't get too close to any object, includingotherboids.
* Alignment:Try to matchthe speed and directionof nearbyboids.
* Cohesion:Head for the perceivedcenterof mass of the boids in your imme-
diate neighborhood.
Reynolds'computationalmethodis called agent-basedmodeling.HadReynolds
chosen insteadto write a top-downprogramfor the global behaviorof the flock,
he might still be workingon it. By choosing insteadto model the flock from the
bottomup, based on agent-levelinteraction,he was able to producehighly realis-
tic flight formationsusing very simple rules.2Note that Reynolds did not model
the flock, nor did he model isolated birds. He modeled their interaction,at the
relationallevel.
Agent-basedmodels (hereafterABMs) of human social interactionare based
on this same theory-buildingstrategy.Sociologists have traditionallyunderstood
social life as a hierarchicalsystem of institutionsand normsthat shape individual
behavior from the top down. Interestin ABMs reflects growing interest in the
possibility that humangroups, like flocks of birds,may be highly complex, non-
linear,path-dependent,and self-organizing.We may be able to understandthese
dynamicsmuch betterby tryingto model them, not at the global level but instead
as emergentpropertiesof local interactionamong adaptiveagents who influence
one anotherin responseto the influencethey receive.
Despite growinginterestin relationalmodelingandcomputationalmethods,so-
ciologists havenot fully appreciatedthe potentialfor ABMs as tools for theoretical
research.This review of recentdevelopmentsis intendedto demonstratehow this
techniquecan providesociologists with a morerigorousmethodfor specifying the
microfoundationsof global patternsat the relationallevel. We begin with a brief
historicalsketchof the shift from "factors"to "actors"in computationalsociology

l"Global" refers to population-leveldynamics that determinethe macro behavior of the

flock, while "local"refers to micro processes at the level of individualsinteractingwith
neighbors.
2Reynold'sboids were so realistic that they provided the startingpoint for bat swarms
in the movies BatmanReturnsand Cliffhanger.You can see the boids in action at www.
1.3.html
discovery.com/area/science/life/life.
 TOACTORS
FROMFACTORS 145

that shows how agent-basedmodeling differs fundamentallyfrom earlier socio-

logical uses of computersimulation.We then review recentcontributionsfocused
on the emergenceof social structureand social orderout of local interaction.
Althoughsociology has lagged behindothersocial sciences in appreciatingthis
new methodology,a distinctive sociological contributionis evident in the papers
we review.First,theoreticalinterestfocuses on dynamicsocial networksthatshape
and are shaped by agent interaction.Second, ABMs are used to performvirtual
experimentsthattest macrosociologicaltheoriesby manipulatingstructuralfactors
such as networktopology, social stratification,or spatialmobility.

HISTORICALDEVELOPMENTOF
AGENT-BASEDMODELS

Computersimulationis more tractable(but less generalizable)thanmathematical

modeling andmore rigorous(butless nuanced)thannaturallanguage(Hanneman
et al. 1995). Gilbert& Troitzsch(1999) identify threeperiodsin the development
of social simulationover the pasthalf-century:macrosimulation,microsimulation,
andagent-basedmodels. In the 1960s, the firstwave of innovationused computers
to simulatecontroland feedbackprocesses in organizations,industries,cities, and
even global populations.With roots in structural-functionalism, macrosimulation
models typically consist of sets of differentialequationsthat predict population
distributionsas a holistic functionof othersystemic factors.Applicationsinclude
the flow of raw materials in a factory, inventorycontrol in a warehouse, state
legitimacy and imperialistpolicy, urbantraffic,migration,disease transmission,
demographicchangesin a worldsystem, andecological limits to growth(Forrester
1971, Meadows et al. 1974, Hannemanet al. 1995).
Beginning in the 1970s, computermodelers introducedthe use of individuals
as the units of analysis but retained the earlier emphasis on empirically based
macro-levelforecasting.In strikingcontrastto the holistic approachin models of
dynamicalsystems, as Caldwell (1997) points out, "microsimulationis a 'bottom-
up' strategyfor modeling the interactingbehavior of decision makers (such as
individuals,families andfirms)within a largersystem. This modeling strategyuti-
lizes dataon representativesamplesof decision makers,along with equationsand
algorithmsrepresentingbehavioralprocesses, to simulate the evolution through
time of each decision maker,and hence of the entirepopulationof decision mak-
ers." Microsimulationmodels use observed populationdistributionsto estimate
parametersfor models of householdcharacteristics(e.g., maritalstatus,numberof
children,labor force status, income, etc.). The models then age the populationat
the individual(or household) level by updatingthese characteristics.Microsimu-
lation models thus resemble the earlier generationof macrosimulationmodels,
but they model changes to each element of the population distributionrather
than changes to the distributionat the population level. However, the models,
do not permitindividualsto directly interactor to adapt.Nor are the models de-
signedor used for basic theoreticalresearch.As with macrosimulation,the primary
146 MACYa WILLER

orientationis towardappliedresearch,mainly forecastingmacroeffects of public

policies thatalterindividualbehavior.
The thirdwave in social simulation,agent-basedmodeling, coincided with the
adventof the personalcomputerin the 1980s. Like microsimulation,these bottom-
up models exploredthe microfoundationsof global patterns.The differenceis that,
unlike the socially isolated actors in microanalyticalsimulation,the agents now
interactinterdependently.More precisely,ABMs impose four key assumptions:
1. Agentsare autonomous.The systemis notdirectlymodeledas a globallyinte-
gratedentity.Systemic patternsemergefromthe bottomup, coordinatednot
by centralizedauthoritiesorinstitutions(althoughthesemayexist as environ-
mental constraints)but by local interactionsamong autonomousdecision-
makers.This process is known as "self-organization"(Kaufman1996).
2. Agentsare interdependent.Interdependencemay involve processes like per-
suasion, sanctioning, and imitation, in which agents influence others in
response to the influence that they receive. Interdependencemay also be
indirect,as when agents' behaviorschange some aspect of the environment,
which in turnaffects the behaviorof otheragents,suchthatthe consequences
of each agent's decisions dependin parton the choices of others.
3. Agentsfollow simple rules. Global complexity does not necessarily reflect
the cognitive complexity of individuals."Humanbeings," Simon contends
(1998, p. 53), "viewed as behaving systems, are quite simple."We follow
rules, in the form of norms,conventions,protocols,moraland social habits,
and heuristics. Although the rules may be quite simple, they can produce
global patternsthat may not be at all obvious and are very difficult to un-
derstand(like Reynolds' "boids").Hence, Simon continues, "the apparent
complexity of our behavioris largely a reflectionof the complexity of the
environment."ABMs explore the simplest set of behavioralassumptions
requiredto generatea macropatternof explanatoryinterest.
4. Agentsare adaptiveand backward-looking.Wheninterdependentagentsare
also adaptive,their interactioncan generate a "complex adaptivesystem"
(Holland 1995, p. 10). Agents adapt by moving, imitating,replicating,or
learning, but not by calculating the most efficient action (Holland 1995,
p. 43). They can adapt at two levels, the individual and the population.
Individualslearnthroughprocesses like reinforcement,Bayesian updating,
or the back-propagationof errorin artificialneuralnetworks.Learningalters
the probabilitydistributionof behaviorscompetingfor attentionwithineach
individual.Populationslearn throughevolutionaryprocesses of selection,
imitation, and social influence. Evolution alters the frequencydistribution
of agent-typescompetingfor reproductionwithin a population.

From Forecasts to Thought Experiments

Unlike earlier approachesto computersimulation,whose value depends largely
on predictive accuracy, agent-based models are "much more concerned with
 FROMFACTORSTO ACTORS 147

theoretical development and explanation than with prediction"(Gilbert 1997,

p. 2.1). They areused to performhighly abstractthoughtexperimentsthatexplore
plausiblemechanismsthatmay underlieobservedpatterns.As such, these models
do not necessarily "... aim to provide an accuraterepresentationof a particular
empiricalapplication.Instead,the goal of agent-basedmodelingis to enrichourun-
derstandingof fundamentalprocessesthatmay appearin a varietyof applications"
(Axelrod 1997, p. 25). When simulationis used to make predictionsor for train-
ing personnel(e.g., flight simulators),the assumptionsneed to be highly realistic,
which usually means they will also be highly complicated."But if the goal is to
deepenourunderstandingof some fundamentalprocess,"Axelrodcontinues,"then
simplicity of the assumptionsis importantand realistic representationof all the
detailsof a particularsettingis not."On the contrary,makingthese modelsmorere-
alisticinevitablyaddscomplexitythatunderminestheirusefulnessas tools fortheo-
reticalresearchif we canno longerfigureouthow themodelproducesa givenresult.
Nevertheless,many sociologists remainhighly skeptical about the validity of
simulationresultswhen computationalmodels areused for theoreticalexploration
ratherthan empiricalprediction.As noted (and lamented)by Sawyer (2001), re-
cent surveyarticleson sociological simulationneglect agent-basedmodeling and
focus primarilyon the earlier equation-basedmethods of social forecasting-
macrosimulationand microsimulation(e.g., Halpin 1999, Hannemanet al. 1995,
Meeker& Leik 1997).
Equation-basedmodels can approximateABMs by operatingon population
attributesand their distribution.The difference, however, is best captured by
Coleman'smethodological"boat"(1990, p. 8). Holistic models representthe deck
of the boat, in which one macrosocialfactor influences another.Coleman advo-
cated a more circuitousroute (via the bottomof the boat), in which initial macro
conditions constrainand motivate the behaviorof individualactors whose inter-
actions then aggregateas a new macrosocialoutcome. This explanatorymethod
searchesfor the causal mechanismsat the level of humanaction thatunderliethe
associationbetween social factors.
ABMs implementColeman'scriticalrealistepistemologybutwithan additional
caveat: The macrosocial outcome is also more than the sum of its parts. This
concept, known as emergence, was anticipatedby Durkheim:"The hardnessof
bronze lies neither in the copper, nor the tin, nor in the lead which have been
used to form it, which are all soft or malleablebodies. The hardnessarises from
the mixing of them" (Durkheim[1901] 1982, pp. 39-40). The principleapplies
as well to sociology, he continued:"[Social] facts reside in the society itself that
producesthem and not in its parts-namely its members."
Here Durkheim oversteps. Whereas the principles of emergence and self-
organizationimply that propertiesof the larger system are not propertiesof the
components-and may not resemble nor be intended by any of the constituent
actors-these principles also incorporatean essential insight of methodological
individualism,the idea that societal patternsemerge from purposivechoices and
not from social facts externalto individuals.Globalpropertiesare sui generis, but
they also emergefromthebottomup, throughlocal interactions.Withouta model of
148 MACY * WILLER

the microfoundationsof emergentproperties,path-dependentself-organizingpro-

cesses (such as informalsocial control) are likely to be mistakenfor institutions
that are globally coordinated(such as bureaucraticcontrols in formal organiza-
tions). In short, ABMs defy classification as either micro or macro but instead
provide a theoreticalbridge between levels (Saam 1999). It is thus ironic that so-
ciological interestin ABMs has lagged behind that of the other social sciences,
for sociology may be the discipline best equippedto develop a methodologythat
bridgesSchumpeter's(1909) methodologicalindividualismand Durkheim'srules
of a nonreductionistmethod.
Clearly,not all problemscan be usefully viewed fromthe bottomup. ABMs are
most appropriatefor studyingprocesses that lack centralcoordination,including
the emergenceof organizationsthat, once established,impose orderfrom the top
down. The models focus on how simple andpredictablelocal interactionsgenerate
familiarbut highly intricateand enigmatic global patterns,such as the diffusion
of information,emergenceof norms,coordinationof conventions,or participation
in collective action. Emergentsocial patternscan also appearunexpectedlyand
thenjust as dramaticallytransformor disappear,as happensin revolutions,market
crashes,fads, and feeding frenzies. ABMs providetheoreticalleveragewhere the
globalpatternsof interestaremorethanthe aggregationof individualattributes,but
at the same time, the emergentpatterncannotbe understoodwithouta bottom-up
dynamicalmodel.
In surveyingrecent applications,we found that most congregatedaroundtwo
problems,(a) the self-organizationof social structureand(b) the emergenceof so-
cial order.The two problemsarehighly complementary.In one case, the clustering
of social ties is the explanandumand in the otherit is the explanans.
1. Emergent structure. In these models, agents and agent-behaviorsmove
throughsocial andphysical space in responseto social influencesand selec-
tion pressures.Agents may startout undifferentiatedand then change loca-
tion or behavior so as to avoid becoming differentor isolated (or in some
cases, overcrowded).Ratherthan producinghomogeneity,however, these
conformistdecisions may aggregateto produce global patternsof cultural
differentiation,stratification,andhomophilousclusteringin social networks.
Otherstudies reversethe process, startingwith a heterogeneouspopulation
andending in convergence:the coordination,diffusion,and suddencollapse
of norms,institutions,beliefs, innovations,standards,etc.
2. Emergentsocial order.These studiesshow how egoistic adaptationcan lead
to successful collective action without either altruismor global (top-down)
imposition of control. A key finding across numerous studies is that the
viability of trust, cooperation,and collective action depends decisively on
the social embeddednessof interaction.
Despite a common focus on two centralproblems,therehas been little effortto
providea meta-analysisof how resultsdiffer dependingon the model designs. To
 FROMFACTORS
TOACTORS 149

thatend, we have groupedstudies by substantiveapplicationin orderto highlight

methodologicaldifferencesthatmay explainconflictingresults.These differences
emerge througha series of interrogationsaboutmodel design:
1. Is interactionglobal or local, that is, is the populationfully connected or is
interactionconstrainedby the structureof social ties?
2. If interactionis local, are the constraintson tie formationspatialor social?
3. Are ties elective (formedandbrokenthroughmovement,exit, or assortative
mating)or is interactionforced?
4. Is adaptationbased on learning(which modifies the probabilitydistribution
of behaviors in each agent's repertoire)or evolution (which modifies the
frequencydistributionof behaviorsacross the populationof agents)?
5. If evolution, does reproductioninvolve competition for survival or social
influence?
6. If influence,is this limited to externalstates of the agent (e.g., behavior)or
do agentscopy otheragents' internalprogramming,even thoughthis cannot
be directlyobserved?
7. Is influence based on attainment(success, fitness, payoffs, status) or fami-
liarity (proximity,frequency)?
8. Is the model used as an experiment(parametersare manipulatedto test for
predicted differences) or a demonstration(parametersare manipulatedto
test for predictedrobustness)?
9. If used experimentally,are the manipulationsmainly of agent-levelparam-
eters (to test a micro theory about the global implications of behavioral
assumptions)or system-level parameters(to test a macro theory about the
dynamicalimplicationsof environmentalassumptions)?
Table 1 classifies representativepaperswe review in a typology based on an-
swers to these nine questions. The articles we included are not intended to be
exhaustive.The field of social simulationis now too large to survey in a single
article. We have thereforenarrowedthe focus to ABMs of emergent structure
(differentiationand diffusion) and emergentorder(cooperationand collective ac-
tion), writtenby sociologists or publishedin sociological journals3in the past five
years.4

3ReviewedarticleswerepublishedinAmericanSociological Review(5), Computationaland

MathematicalOrganizationTheory(5), JournalofArtificialSocieties and Social Simulation
(5), AmericanJournal of Sociology (3), Rationalityand Society (2), Sociological Methods
and Research (1) andBehavioralScience (1).
4ForanAnnualReviewof relatedpaperspublishedbefore 1996, see Bainbridgeet al. (1994).
For a recentreview of agent-basedmodeling in political science, see Cederman(2001) and
Johnson(1999).
150 MACY ? WILLER

og t
I I

00

. o oa o -

CoCo o

IM o 9C
0 .I o

CC .00. C
Cge F3 03 2o
0 0 o

aa JO

Co 4 C?

0 0 0
O o
00

Ct U0
caca ? 'ca 0 ? 0 C
QQQQQe

8
C -0 " v,
o P~
!o
ooo
o

a
4-

C C .&
 FROMFACTORSTO ACTORS 151

EMERGENT STRUCTURE: MODELS OF CONVERGENCE

AND DIFFERENTIATION

In models of structuraldifferentiation,interestcenters on the self-organizationof

the populationinto locally dense networks(or clusters) based on simple rules of
local interaction.Applicationsinclude residentialsegregation,density-dependent
organizationalsurvival,groupformation,and culturaldifferentiation.
These models often studyclusteringwithin spatialnetworks,using cellularau-
tomata(CA), a techniquefirstproposedby StanislawUlam (Coveney& Highfield
1995, pp. 94-96). Hegselmann& Flache (1998) provide a lucid introductionto
andhistoryof CA in the social sciences. The agentsusuallylive on a checkerboard
(eitherflat or a donut-liketorus),andthe state of each agent dependson the states
of its neighbors.Simple rules of local influence or spatialmovement sometimes
generatesurprisingresults and lead to unexpectedinsights. They illustratea key
advantageof the CA approach:two-dimensionalvisual representationof diffusion
and clusteringacross a spatialnetwork.
Schelling's (1971) model of neighborhoodsegregationis one of the earliest
and best known ABMs based on movement in a spatialnetwork.Red and green
agents are randomlydistributedon a lattice and move to empty locations if the
numberof in-groupneighborsfalls below a certainthreshold.The model shows
how extremesegregationtends to arise even in a populationthatprefersdiversity,
as agents relocateto avoid being in the minority.
One criticism of many spatial networks is that they preclude both structural
equivalence (no two nodes can have identical sets of interactants)and relational
heterogeneity (every node has an isomorphic relational pattern). Flache &
Hegselmann(2001) relaxedthe latterconstraintby using irregulargridsthatallow
the numberand strengthof social ties to varyrandomlyover the population.How-
ever, the model still permitsthe populationto self-organizeinto clusters through
spatialmovement.Althoughthey find thatkey results are robustover variationin
relationalheterogeneity,irregulargrids also yield new implicationsthatcould not
be identifiedwith the conventionalrectangularstructures.
Other CA models generate spatial dynamics through ecological5 competi-
tion for survival ratherthan movement. The classic in this genre is Conway's
"Gameof Life" in which the survivalof each agent dependson the density of its
neighborhood.6Although Conway was not a sociologist, his design has imme-
diate applicationto problems in organizationalecology, in which the agents are

SEcologicalmodels arethose in which the spatialorfrequencydistributionof agentsdepends

on rules that govern survival and reproduction.These models are often characterizedas
"evolutionary,"but strictly speaking, the latter requires the possibility for entirely new
types of agents to appearthat were not presentat the outset.
6LikeSchelling's, Conway'smodel was originallycreatedwithouta computer,using a game
board.The model is now implementedin Javaandinterestedreaderscan experimentonline
at www.math.com/students/wonders/life/life.html. However,these simple CA models are
not agentbased as defined above.
152 MACYm WILLER

supra-individual.For example, Lomi & Larsen (1998) study the interactionbe-

tween networkstructureand the lagged effects of populationdensity on organi-
zational survival.They use a cellular networkin which the survivalof each cell
depends on the numberof occupied cells in its Moore neighborhood(the eight
adjacentcells). The model is very simple:Agents live, replicate,anddie, based on
local density.Lomi & Larsenthen explorethe implicationsfor organizationalsur-
vival of alternativehypothesesaboutthe effects of density delay based on simple
rulesthatregulatethe appearance,survival,anddemise of individualorganizations.
Using event historyanalysis, they identify structuralfeaturesthatcan generateor-
ganizationallife histories that are qualitativelyconsistent with those observedin
empiricalorganizationalpopulations.
Some models incorporateboth spatialmovementand ecological competition.
In Epstein & Axtell's (1996) "Sugarscape,"a spatiallydistributedpopulationof
simple rule-basedagents develops a culture, an economy, and a class structure.
Agents move aroundon a grid and exchange with others to gain access to valued
resourceson which their survivaland reproductiondepend.

Social Influence and the Paradox of Mimetic Divergence

The ecological assumptionthat adaptationoccurs througha strugglefor survival
is appropriateif the agents areorganizationscompetingfor resourcesor members.
If the agents are individualsin a modem welfare state, however,a more broadly
applicable assumptionis that adaptationoccurs throughimitation of the fittest.
Agents arenot replacedby betterperformers;they simply copy theirobservedbe-
havior.This in turnrelaxesthe assumptionthatselectionpressuresareperformance
driven.Although some influencemodels continue to posit selection of role mod-
els based on relative success, others assume that influence is density dependent,
basedon familiarity,popularity,or spatialproximity.Forexample,Latane's(1996)
Social ImpactModel uses a rule to mimic one's neighborsin a two-dimensional
lattice.Froma randomstart,a populationof mimics mightbe expectedto converge
inexorablyon a single profile, leading to the conclusion that culturaldiversityis
imposed by factorsthatcounteractthe effects of conformisttendencies.However,
the surprisingresult was that"the system achieved stable diversity.The minority
was able to survive,contraryto the belief thatsocial influenceinexorablyleads to
uniformity"(Latan61996, p. 294).
Following an earlier study by Carley (1991), Axelrod (1997; see also Axtell
et al. 1996) makes the paradox of mimetic divergence even more compelling.
Carley'sandAxelrod'smodels couple local influence(thetendencyfor people who
interactfrequentlyto become moresimilarovertime) andhomophily(the tendency
to interactmore frequentlywith similar agents). This closes the loop-the more
agents interact,the more similarthey become, and the more similarthey become,
the more likely they are to interact.More precisely,neighboringagents on a two-
dimensionallattice interactwith a likelihooddeterminedby the similarityof their
culturaltraits(given by a simple, randomlyassigned stringof numbers).Interac-
tion, in turn,reducesremainingdifferences.Axelrodexpectedthis self-reinforcing
 FROMFACTORSTO ACTORS 153

dynamic would lead inexorablyto culturalconvergenceand homogeneity.Again

the result was surprising.He found that "local convergence can lead to global
polarization"and that unique subculturescan survive in the face of a seemingly
relentless march towardculturalconformity.Stable minority subculturespersist
because of the protectionof structuralholes createdby culturaldifferences that
precludeinteraction,therebyinsulatingagents from homogenizing tendencies.
Axelrod's model also reveals a surprisingeffect of populationsize. Intuitively,
one might expect largernumbersof stable subculturesto emerge in largerpop-
ulations. However, Axelrod found a nonlinear effect, in which the number of
minority cultures first increases with population size but then decreases. This
counter-intuitiveresultillustratesthe principleof "gambler'sruin."Largepopula-
tions allow for largerculturalmovementsthat can surviverandomfluctuationsin
membershipbetterthansmallercompetitors.As the big get bigger,the numberof
minoritysubculturesdiminishes.
Axelrod begins with a heterogeneouspopulationand shows thatheterogeneity
persists.But how does the initialheterogeneityarise?Axelrodalso assumesspatial
networksthat restrictinteractionto nearbyneighbors.Will differentiationpersist
if the spatialrestrictionis removedand interactionis based only on similarity?
Mark (1998) addressesthese questions in a paper that explains social differ-
entiation"from first principles,"startingfrom homogeneity and without spatial
constraintson interaction.Agents can interactwith anyonein the population(not
just neighbors),with a probabilitydeterminedby culturalsimilarity.Startingfrom
perfect culturalhomogeneity,interactionsare initially random,but not for long.
Mark finds that a self-reinforcingdynamic based on homophily and creation of
new bits of culture,is sufficientto createan emergentnetworkwith local patterns
of interactionamongdistinctivesubcultures.Contraryto Axelrod,Markalso finds
thatpopulationsize decreasesculturalhomogeneity,due to the absence of spatial
restrictionson interactionin the model.
One limitationin most social influencemodels is the assumptionthatinfluence
is only positive. However,social relationscan also havenegativevalence, such that
the state of an agent tends towardmaximal distinctivenessratherthan similarity.
Contraryto theoriesof homophily,dissimilaritydoes not always weakenthe social
tie; rather,it may sometimesstrengthenthe negativerelation(orenmity).Structural
differentiationbased on positive and negative influence has been studied using
attractorneuralnetworks,a cognitive modeling techniquedevelopedby Hopfield
(1982) andappliedto social influenceby Nowak & Vallacher(1998; see also Kitts,
et al. 1999).
Artificial neural networks are a simple type of self-programmablelearning
device basedon paralleldistributedprocessing(Rummelhart& McClelland1988)
and modeled after the nerve systems of living organisms. In elementary form,
the device consists of a web of neuron-likeunits (or neurodes) that fire when
triggeredby impulses of sufficientstrength,and in turnstimulateor inhibit other
units when fired.The effect of an impulse (as stimulusor inhibitor)dependson the
sign and strengthof the synapticconnectionbetween two neurodes.The network
learnsby modifying these pathcoefficients.In feed-forwardnetworks,learningis
154 MACY* WILLER

based on environmentalfeedbackthatpropagatesbackwardthroughthe network.

In attractornetworks, the path coefficients are updatedbased on the similarity
betweenthe statesof adjacentconnectednodes. While feed-forwardnetworkscan
be used to model agentcognition,attractornetworksprovidea dynamicalternative
to static networkmodels of social interaction.
Feed-forwarddevices have been used to model cognitive social differentia-
tion, based on self-affirmingstereotypes.Vakas-Duong& Reilley (1995; see also
Bainbridge1995) study the emergenceof irrationalracial hiringpreferencesthat
are less profitablethan purely meritocraticselection. In their model, employers
learn to make intuitivehiring decisions based on what connotationscome to be
associated with the traitsexhibited by job applicants,while applicantsassociate
traitswith relative success. The simulationresults showed how false beliefs can
easily become self-sustainingfollowing an early accidentof associationthatsows
the seeds of racialpreferencein an employer'smind.This in turnmakesit difficult
for talentedmembersof the same race to gain employmentand diminishes that
race's access to emergingstatus symbols.
Other cognitive models of social differentiationfocus on the self-reinforcing
dynamics created by stereotypicalbeliefs that change the behaviors on which
the beliefs are based. Orbell et al. (1996) model self-organizingstereotypesin a
populationof 1000 adaptiveagents playing Prisoner'sDilemma games with an
option to exit. Prisoner'sDilemma is a two-persongame in which the best move
is to defect (e.g., cheat), no matterwhat the partneris expected to do, but when
each defects, the outcome is deficient for both. However, the authors' interest
is focused not on the problem of cooperationbut the formationof groups and
group stereotypes.Agents are assigned a tag that indicates their membershipin
one of two groups. Agents update their tag-specific expectationsof cooperative
behaviorbasedon the outcomesof interactions.Agents also become morelikely to
cooperatewith membersof groupsthey expect to cooperate.They findthatagents
frombothgroupsconvergeon anincreasinglystrongpreferenceforinteractionwith
membersof an initiallypreferredgroup(when groupsare of equal size) or a larger
group (when initial preferences are equal). The mechanism is straightforward:
If you interactwith one group more than the other (due either to differences in
initial expectations or group size), you update your expectationsfor this group
more thanfor the other.Since both groupsarerelativelycooperative,the updating
alwayscauses expectationsof cooperationto increase,leadingto morecooperative
treatment,which inducesthe expectedcooperativebehavior,therebyaffirmingthe
expectationon which the preferentialinteractionis based.

Diffusion of Innovation
The models consideredso far explore emergentnetworksbased on structuraldif-
ferentiation.However, social influence models can also be used to study self-
reinforcingdynamicsthatlead to convergence.Applicationsincludethe spreadof
innovations,coordinationof conventions,emergentnorms,and culturaldiffusion.
 FROMFACTORS
TOACTORS 155

These models startwith some distributionof practicesand a rule by which agents

decide whetherto abandoncurrentpracticein favorof one used by anotheragent.
Rosenkopf& Abrahamson(1999) studieddiffusionwhere influencederives from
popularity,without regardto prior interactionhistory or relative success. This
creates a "positive feedback loop where adoptions by some actors increase the
pressure to adopt for other actors" (Rosenkopf & Abrahamson1999, p. 361).
However, influence was weighted by reputations(which were exogenous to the
model) and combined with informationabout the unprofitabilityof innovations.
The networkwas fully connected,that is, each agent had access to the decisions
and reputationsof all other agents in the population. They found that "band-
wagons occur even when potential adopters receive informationabout others'
unprofitableexperienceswith the innovation"(1999, p. 361). Their model shows
how agentscan convergeon inefficientpracticesbutnot how conformitycollapses.
Bullnheimeret al. (1998) studiedthe diffusionof innovationamongfirmsbased
on the assumptionthatmanagersevaluatethe relativeperformanceof alternative
technologies based on their own experiencewith a technology as well as with the
performanceof others. They assume that firmshave knowledge of the adoptions
and performanceof all otherfirms.They found that firmswho combine imitation
with their own experience outperform"both pure imitatorsand nonimitatorsin
productionefficiency as well as in profits"(1998, p. 267). Like Rosenkopf &
Abrahamson,they explainstablehomogeneitybutnot how conformitymight then
collapse.
In contrast,Strang & Macy (2001) generate the punctuatedcascades associ-
ated with fad-likebehavior.Theirdiffusionmodel bridgesthese earlierstudies by
showinghow a decision rule similarto those Bullnheimeret al. found to be highly
efficientcan neverthelesstrapfirmsin a bandwagonof adoptionandabandonment
of innovationsthatare worthlessor nearlyso. They assume firmsevaluatecurrent
practicebased on their balance sheets, and if dissatisfied,turnto "bestpractices"
for new ideas. In a series of computationalexperiments,they manipulatethe in-
trinsicvalue of innovations,the stratificationof the market,and the skepticismof
managersto see how these affect the fad-like pattern.Results show that fads are
most likely in stratifiedmarketswhere innovationshave a modest effect on per-
formanceand managersare not so skepticalthatthey cannot see the performance
differences.

EMERGENT ORDER: MODELS OF COLLECTIVE

ACTION, TRUST, AND COOPERATION

In models of structuraldifferentiation,agents influence othersin response to the

influence they receive, leading to spatial or social clustering, such as Reynold's
flocks of "boids."Interestcenters on the self-organizationof dynamic structural
configurations,and not on theirconsequences.Models of emergentorder,in con-
trast,focus attentionon the ways in which networkstructuresaffect the viability of
156 MACYE WILLER

prosocial behavior.Fournetworkpropertieshave been shown to promote/inhibit

cooperationand participationin collective action:
* Relational stability:On-going relationshipslengthenthe "shadowof the fu-
ture"(Axelrod 1984).
* Networkdensity:The coordinationcomplexityof cooperationincreaseswith
the numberof social ties (Macy & Skvoretz 1998).
* Homophily:Agents tend to interactwith partnerswho use similarstrategies
(Cohen et al. 2001).
* Transitivity:An agent'spartnerstend to interactwith each other.This in turn
affects:
* Diffusion of reputations(Takahashi2000).
* Bandwagonscaused by thresholdeffects (Chwe 1999).
* Monitoringand enforcementof conformity to prosocial norms (Kim &
Bearnan 1997).

RelationalStability
The classic studyof emergentorderis Axelrod's (1984) Evolutionof Cooperation.
Althoughdefectionis the dominantstrategyin a single play of Prisoner'sDilemma,
that is not true when the game is played repeatedlyin an ongoing relationship.
However,thisdoes not guaranteecooperation.In fact, thereis no dominantstrategy
in endlessly repeatedplay, and game theorycannotpredictwhethercooperationis
likely to evolve. To find out, Axelrod organizeda computertournamentin which
agentsplayed a roundrobiniteratedPrisoner'sDilemma. He invitedleading game
theorists to submit strategies,and each submission was assigned to one of the
agents. The winner was the simplest contestant,Anatol Rapaport'sTit for Tat, a
strategythat always cooperates unless provoked,and then always retaliates,but
only once and then forgives.
Variationson Tit for Tatuse less-strict accounting.De Vos et al. (2001) com-
parethe evolutionaryviability of two types of reciprocity,based on strictvs. loose
accounting.Payoffs determinewhetheran agentsurvives,andstrategiesreproduce
in proportionto the survivalrate.Agents ask for help when it is needed anddecide
to whom to give it (if anyone), based on past exchanges. Some agents insist on
keepingthe books strictlybalanced,while othersfavorcommitmentto old partners,
even if they are in arrears.In a populationthat includes nongivers, simulations
demonstratethe importanceof committing oneself to an ongoing relationship.
Ironically,the results suggest that a looser commitmentstrategy,based on long-
term balancingof the books, is superiorto a strategyof strictreciprocitythat is
less vulnerableto being cheated,a resultsimilarto thatreportedby Kollock (1993)
based on a similarecological competition.However,Kollock's agentswere paired
randomly,withoutthe option to select theirpartners.He also found that loose ac-
countingis superior,butonly if the environmentis noisy (with occasionalmistakes
 FROMFACTORSTO ACTORS 157

andmisinformation).Strictreciprocityis thenproneto needless recriminationthat

can be avoidedby looser accountingsystems.
One problemwith this modeling strategyis thatthe outcomeof an evolutionary
tournamentmay be an artifactof a theoreticallyarbitraryset of initial contestants.
This led Axelrod to use a genetic algorithm(GA) to see if Tit for Tatwould evolve
in an open-endedpopulationin which strategiescould evolve from a randomstart
(1997, pp. 14-29). Workingwith JohnHolland,Axelrod found severalstrategies
similarto Tit for Tatthatprovedto be highly robust.
GAs are stringsof computercode that can mate with other stringsto produce
entirelynew and superiorprogramsby buildingon partialsolutions.Each strategy
in a populationconsists of a stringof symbols that code behavioralinstructions.
These symbols are often binary digits (or bits) with values of 0 or 1. A string
of symbols is analogous to a chromosome containing multiple genes. A set of
one or more bits thatcontains a specific instructionis analogousto a single gene.
The values of the bits and bit combinationsare analogous to the alleles of the
gene. A gene's instructions,when followed, producean outcome (or payoff) that
affects the agent's reproductivefitness relative to other players in the computa-
tional ecology. Relative fitness determinesthe probabilitythat each strategywill
propagate.Propagationoccurs when two mated strategiesrecombine.If two dif-
ferent rules are both effective, but in differentways, recombinationallows them
to create an entirely new strategy that may integratethe best abilities of each
"parent,"making the new strategy superiorto either contributor.If so, then the
new rule may go on to eventuallydisplace both parentrules in the populationof
strategies.In addition,the new stringsmay containrandomcopying errors.These
mutationsrestorethe heterogeneityof the population,counteractingselectionpres-
suresthattend to reduceit.
The GA can be used to discover both optimal and likely solutions. Where
the aim is to discoverwhat agents should do to optimize performance,the models
typicallyassumeglobal search.This meanseveryagenthas completeknowledgeof
the strategiesandfitnessof everymemberof the populationandplays againstevery
memberof the populationwith equal probability.Where the aim is to find what
agents are likely to do, models often assume local ratherthan global interaction
and knowledge (Klos 1999). Local searchcan be implementedby embeddingthe
GA in a spatialor social network.

Network Density
Macy & Skvoretz(1998) embedthe GA in a social networkto investigateWeber's
theorythatProtestantsects in colonial Americaprovidedculturalmarkersneeded
for trusting strangersin physically dispersed markets.The problem in Weber's
argumentis thatthe need for economic growthdoes not guaranteethe evolutionof
the means for its realization.Macy & Skvoretz's simulationsshow that a system
of telltale signs is highly fragile, even with unrealisticallygenerous assumptions
about culturaldiffusion. However,the robustnesscan be greatly improvedwhen
158 MACY* WILLER

exchangesareembeddedin social structurescomprisedof a largenumberof small

communities,precisely the conditionsthatWeberidentifiedin colonial America.

Homophily
Smith& Stevens(1999) model the formationof psychologicalsupportnetworksin
which agentsseek outrelationshipswith othersthatwill help themmanageanxiety.
In their model, agents decide with whom to form relationshipsthrougha process
of assortativemating.They find that agents form relationshipswith partnerswho
are similarto themselves in their ability to manage stress, creatinghomophilous
clusters.In needy populations,supportnetworksform with strongerattachments
but lower transitivitythanin populationswith less need for social support.
Several other recent studies also suggest that the viability of cooperationis
greatly improvedwhen populationscan self-organizeinto locally homogeneous
clusters (Lomborg 1996). For example, Pedone & Parisi (1997) use socially em-
bedded artificialneuralnetworksto show how altruisticbehaviorcan arise among
similaragents,andthey concludethatsimilarityconveyedby culturemay be what
allows altruismto evolve in naturalsettings.
Otherstudiesexplorethe effects of homophilousclusteringin spatialnetworks.
Eshel et al. (2000) use spatialclusteringon a one-dimensionalarraywhere agents
play Prisoner'sDilemma.Theiragentshave only two possible strategies-coope-
rateor defect. Agents interactstrategicallywithnearestneighborsandimitatethose
(in a somewhatlargerneighborhood)who aremost successful.Becausetheycannot
reciprocate,thereis no advantageto cooperatingeven in an ongoing relationship.
Nevertheless,when the game is spatiallyembedded,they find that cooperationis
"a stable strategythat cannot easily be eliminated from the population"(2000,
p. 341).
Flache & Hegselmann (1999) explore the macro implications of alternative
assumptionsabout agent cognition in a social supportgame played on a torus.
Rational agents make the choices prescribedby analytical game theory, while
adaptive agents respond to experience throughreinforcementlearning. In each
case, agents migrateon the grid, selecting neighborsfrom whom to requesthelp
and deciding whether to service the requests of others in an asymmetricPris-
oner'sDilemma, where payoff asymmetryreflectsdifferenceamong agents in the
need for andabilityto help. The forward-lookingmodel makesstrongassumptions
aboutinformation:Eachagentknows all players' locations,theirpayoffs, andtheir
level of need. Flache& Hegselmannfindthatboth forward-andbackward-looking
agents self-organizeinto mutuallysupportiverelationshipswith those in similar
class positions.However,the stratificationstructurediffersundermost conditions.
Rational egoists tend to form an "onion-shapedstructureof solidarity"(1999,
p. 110), with the wealthiest at the center of a large heterogeneouscluster, sur-
rounded by rings of increasingly needy agents. In contrast, backward-looking
agents migrateinto distincthomogeneousclusterswith much greaterclass segre-
gation thanfound among rationalegoists.
 FROMFACTORS
TOACTORS 159

In sum, numerousstudies converge on the conclusion that prosocial strategies

thriveon both spatialandsocial embeddedness,due to the tendencyto interactwith
similar strategies while avoiding contact with predators.However, Cohen et al.
(2001) point out thatwhat appearsto be the effect of homophilymay actuallybe
due to the effects of relationalstabilityand transitivity.Transitivity(or clustering)
means that "pairedagents have neighbors who are themselves paired"(Cohen
et al. 2001, p. 11). Forexample,in a Mooreneighborhood,each of an agent'seight
neighborsinteractswith two of the other seven. Relationalstability (which they
call "contextpreservation")means that agents continueto interactwith the same
partnersacross many periods, creatinga "shadowof the adaptivefuture"(2001,
p. 13). The smaller the neighborhood,the greaterthe chance of interactingwith
a previous partner.In short, the effect of embeddednessmay not be due to the
tendencyfor local interactionsto be with partnerswho are similar,as most studies
have assumed, but the tendency for partnersto be correlated(due to network
transitivity)or familiar(due to local pairing).
The authorstease apartthese effects by manipulatingnetwork structurein a
population of 256 agents who play a four-iterationPrisoner's Dilemma game
with each of four differentpartnersin each period, for 2500 periods. Agents are
programmedwith three(initiallyrandom)probabilitiesfor cooperatingunderthree
conditions: the first move of the game, after the partnercooperates, and after
the partnerdefects. Between periods, agents adoptthe strategyused by the most
successful of their four partners,based on the payoffs accumulatedduring that
period. In this model, many different strategiescan produce identical behavior
under given conditions, and the authorsdo not explain how agents know which
of these was actuallyresponsiblefor the behaviorthey observe, but somehow the
agents guess correctlyabout90% of the time.
Using controlledcomputationalexperiments,the authorsobserve the indepen-
dent effects of homophily,ongoing relations, and transitivity.They find that on-
going relationsgreatly improvethe viability of cooperation,whereastransitivity
alone has little effect. However,the effect of relationalcontinuityis not due to the
prudenceof being nice to those one expects to meet again. Because agents imitate
their partners,ongoing relations increase the chances that an agent will interact
with a partnerusing a similar(if not identical)strategy.They concludethatfriendly
strategiesdo well so long as they can generallyavoid those that are not.

Diffusion of Reputations
Takahashi(2000) uses an evolutionarymodel to study the emergence of gener-
alized exchange, in which agents give and receive help but not to one another
directly (see also Nowak & Sigmund 1998). Takahashichallenges previous stud-
ies thatassumedthatthese exchangesystems requireeitheraltruismor centralized
enforcementof the rules of exchange.He then uses an evolutionarymodel to show
thatexchangesystems can self-organizebasedon normsof generalizedreciprocity
(giving selectivelyto those who give to thirdparties).He programsagentswith two
160 MACY? WILLER

genes that control compliancewith norms of generalizedexchange and enforce-

ment of complianceby others.The firstgene controlsthe amountthe agent gives
to others, and the second gene controlsreciprocity,based on the recipient'srepu-
tationfor giving to others. Giving and receiving determineseach agent's relative
fitness or chances for reproduction.Reproductioncopies the agent's genes with a
small probabilityof mutation.However,with only two genes, thereis no need for
recombination,so Takahashidoes not use a genetic algorithm.Simulationsshow
that a system of generalizedexchange can evolve in a populationthat is initially
nongenerous,assuming agents have perfect informationabout the past behavior
of other agents. Takahashithen relaxes this assumptionby positioning agents on
a two-dimensionalgrid, restrictingtheirknowledge,interaction,and reproductive
competitionto their Moore neighborhood.Thus, agents continue to have perfect
informationabout all their potential exchange partners,of whom there are now
only 8 (insteadof 19). Generalizedexchangeemerges within each of the overlap-
ping neighborhoods,butTakahashididnot test to see if generalizedexchangecould
evolve betweenmembersof differentneighborhoodswhenreputationalknowledge
remainslocal.
Castelfranchiet al. (1998; see also Conte & Castelfranchi1995) examine the
effect of reputationson deterrenceof aggressive behavioron a two-dimensional
grid where agents compete locally for scarce resources and adaptationoperates
throughevolutionaryselection. They find that a prosocial strategycan thrivein a
homogeneouspopulationbut suffersas contactwith aggressorsis increased.How-
ever, the aggressor'sadvantageis diminishedif agents can exchange information
on the reputationsof others.Saam& Harrer(1999) used the same model to explore
the interactionbetweennormativecontrolandpower.They findthatsystems of in-
formalsocial controlcan tip towardeithergreaterequalityor inequality,depending
on the extent of inequalityat the outset.

Bandwagons
Networktransitivitybecomes muchmoreimportantwhen outcomesdependon the
flow of informationthroughthe network.Chwe (1999) proposesa thresholdmodel
of collective actionin which agentschoose to participatedependingon the number
of neighborsexpected to participate.Expectationsof neighbors'behaviordepend
in turnon expectationsof neighbors'neighbors'behavior,and so on. In the base
condition, 30 agents are randomlyassigned two partnerswith whom they remain
attachedfor the durationof the simulation.Chwe then manipulatestransitivityby
increasingthenumberof partnersandthebias towardselectingthepartnersof one's
partners.High transitivityavoids an endless regressbecause an agent's neighbors
and the neighbors' neighbors are likely to be the same people. Transitivityis
especially importantin populationswith low thresholdsthat can be triggeredby
local knowledge about the behavior of members of densely tied but relatively
smalllocal clusters.This may explainthe importanceof overlappingsocial ties for
Freedom Summerparticipationreportedby McAdam (1988). Conversely,Chwe
demonstratesthe strengthof weak ties in populationswith high thresholds.Low
transitivityfacilitates the diffusion of informationabout participationof distant
 FROMFACTORSTO ACTORS 161

agents. The optimal configurationmay be a small worlds network(Watts 1999),

with a few ties between many small and densely tied clusters.

Social Pressure
Chwe's thresholdscorrespondto agents' concernsaboutthe efficacy of participa-
tion in collective action (see also Marwell& Oliver 1993, Macy 1991). Thresholds
can also representagents' responsiveness to social pressures to conform to an
emergentnorm, as in bandwagonmodels of self-reinforcingpopularity.Kim &
Bearman(1997) model collective actionamongagentswhose interestin the public
good is heavily influencedby social pressurefrom otherparticipantsin theirlocal
network. This causes interest in the collective action to spreadlike a contagion
throughnetworkchannels.The authorsfind thatparticipationspreadsmost effec-
tively within densely clusteredsubnetworkscomprisinga critical mass of highly
interestedprimarycontributors.
Kim & Bearman'sstudy reflects the conventionalwisdom that social pressure
to participateis needed to overcome the temptationto free ride. It follows that
dependence on the group for social direction promotes compliance with group
obligations, as argued by Homans (1974). Yet a number of ethnographicstud-
ies of deviant cliques have shown that conformistpressurescan also undermine
normativecompliance,leading to badly suboptimaloutcomes for all groupmem-
bers, including the deviants (Willis 1977, Shibutani1978, MacLeod 1995). This
led Flache & Macy (1996) to investigatethe possibility that dependenceon peer
approvalcan backfire, leading to collective action failure ratherthan success.
Kitts et al. (1999) extended this study by modeling self-organizingsocial rela-
tions using an attractorneural network similar to Nowak & Vallacher's(1998).
In these models of dynamicnetworks,social influenceincreaseswith the strength
of the connectingtie, and ties strengthenwith the similarityof connected agents.
They added the innovationthat agents respond not only to social influence but
also to the lessons of direct experience (similarto the back-propagationin feed-
forwardneural nets). Agents in a team-rewardedtask group decided whetherto
work or shirk and whetherto approveof othergroup members.The authorsthen
manipulatedagents' susceptibilityto influence. Computersimulations revealed
a surprisingresult-a curvilineareffect of social influence on compliance with
prosocial norms. Moderatedoses of influence reduce the coordinationcomplex-
ity of self-organizedcollective action and help the network achieve satisfactory
levels of cooperation.High doses, however,underminethe agent-basedlearning
requiredto findcooperativesolutions.Increasinggroupsize also diminishedcom-
pliance due to increasedcomplexity,with largergroupsrequiringmore influence
to overcome the coordinationproblem.

CONCLUSION

Agent-basedmodeling is a new tool for theoreticalresearchat the relationallevel,

with particularrelevancefor sociologists as a bridgebetweenthe micro andmacro
levels. Nevertheless, sociology has lagged behind the other social sciences in
162 MACY? WILLER

recognizingand exploiting this methodology.Computationalsociology has tradi-

tionally used simulationto forecast social trajectoriesbased on statisticalassoci-
ations, using models that are highly realistic, empiricallygrounded,and holistic.
In contrast,agent-basedmodels use simulationto search for causal mechanisms
that may underlie statistical associations, using models that are highly abstract
and microsocial. This has led to confusion about the appropriatestandardsfor
constructingand evaluatingagent-basedmodels. We conclude our review with
a series of recommendationsfor realizing the rich sociological potential of this
approach.
1. Start it simple. Analysis of very simple and unrealisticmodels can reveal
new theoreticalideas thathave broadapplicability,beyondthe stylized mod-
els thatproducedthem. Pressureto make models morerealistic (and agents
more cognitively sophisticated)is misguided if models become so com-
plex that they are as difficult to interpretas naturalphenomena. When
researchersmust resort to higher order statistical methods to tease apart
the underlyingcausal processes, the value of simulationis largely under-
mined. Models should startout simple and complicationsshould be added
one at a time, making sure that the dynamics are fully understoodbefore
proceeding.
2. Avoidreliance on biological metaphors.Chattoe(1998) has raised probing
questions about modeling culturalevolution as a genetic analog. What is
the mechanism that eliminates poor performersfrom the population and
allows others to propagate?Imitationof the fittestmay be more applicable
than starvationand reproduction,but unlike survivalof the fittest,mimetic
selectionreplicatesonly observedbehavioranddoes not copy the underlying
(unobservable)rules. Biological metaphorspaper over the importanceof
this distinction.For example, in repeatedPrisoner'sDilemma games, it is
plausiblethatan agent observes and then copies the cooperativebehaviorof
successfulneighbors,buthow does the agentknowthatthis behavioris based
on a rule (or supergamestrategy)like Tit for Tat and not Win-Stay,Lose-
Shift or Always Cooperatethatgenerateidenticalbehaviorwith cooperative
partners?
3. Experiment,don't just explore. Agent-based modeling is an experimental
tool for theoreticalresearch.While importantdiscoveries can be made by
open-endedexplorationof theoreticalpossibilities, researchersneed to re-
sist the temptationto become freewheelingadventurersin artificialworlds.
Careful, systematic mapping of a parameterspace may be less engaging,
but it makes for betterscience. This requirestheoreticallymotivatedmani-
pulation of parameters,based on careful review of currenttheoreticaland
empiricalknowledge, and a clear statementof the hypotheses that guided
the experimentaldesign.
4. Testrobustness.Althoughsimulationdesigns shoulduse experimentalrather
thanpost-hocstatisticalcontrolsto identifyunderlyingcausalprocesses,that
 FROMFACTORSTO ACTORS 163

does not mean researchersshould avoid statisticalanalysis of the results.

On the contrary,ABMs, especially those that include stochasticalgorithms,
requirereplicationsthatdemonstratethe stabilityof theresults.Wherepossi-
ble, replicationsshouldincludevariationin parametersthataretheoretically
arbitraryor of secondaryinterest. Authors then need to be careful to dis-
tinguishbetween experimentalmanipulations(whereresultsareexpectedto
change with the parameters)and robustnesstests (where they are not).
5. Testexternalvalidity.Virtualexperimentstest the internalvalidity of a the-
ory, without which there is no need to test the externalvalidity. However,
this does not mean thereis never such a need. ABMs are often used to grow
familiar macrosocial patterns,as a way to identify possible causal mech-
anisms (Epstein & Axtell 1996). When this succeeds, researchersneed to
think about ways these mechanisms can be operationalizedand tested in
laboratoryor naturalconditions.
6. Testdomainvalidity.Do two differentmodels of the samephenomenonpro-
duce the same results?If they do not, findout why. Identifythe assumptions
that are the logical source of divergentimplications by carefully aligning
the models. "Withoutsuch a process of close comparison,computational
modeling will never provide the clear sense of 'domain of validity' that
typically can be obtained for mathematizedtheories" (Axtell et al. 1996,
p. 123).
7. Bring factors back in. The bottom-up approachin ABMs might seem to
imply thatthese models can only be used to test microsociologicaltheories.
That is a tragic misunderstandingbecause it precludeswhat is most excit-
ing to sociologists about this methodology. An artificialworld populated
by computationalagents is a laboratoryin which researcherscan manipu-
late structuralconditionsto test macrosociologicaltheorieswithoutreifying
causal factorsat the macrolevel. Contraryto the holistic epistemology of an
earliergenerationof equation-basedsimulations,changesin populationden-
sity or networkstructure,for example, do not directly lead to the diffusion
of innovations.The causal process is effected throughindividualchoices.
Computationalexperimentsin virtualworlds provide a rigorousmethodol-
ogy for studying the microfoundationsof macro dynamics. However, the
shift from factors to actors should not limit experimentersto manipulation
of agent attributes(such as cognitive or behavioralassumptions).Bringing
factorsback in as experimentalmanipulationswill realize the full potential
of agent-basedmodeling, especially in sociology.

ACKNOWLEDGMENTS
The first authorexpresses gratitudeto the National Science Foundation(SES-
0079381) for supportduringthe period in which this review was written.We also
thank James Kitts, Andreas Flache, and Noah Mark for helpful comments and
suggestions.
164 MACYa WILLER

The AnnualReviewof Sociologyis online at http://soc.annualreviews.org

LITERATURE CITED
Axelrod R. 1984. The Evolution of Coopera- ColemanJS. 1990. Foundationsof Social The-
tion. New York:Basic ory. Cambridge,MA: HarvardUniv. Press
Axelrod R. 1997. The Complexityof Coopera- ConteR, CastelfranchiC. 1995. Understanding
tion. Princeton,NJ: PrincetonUniv. Press the effects of normsin social groupsthrough
Axtel R, Axelrod R, Epstein JM, Cohen MD. simulation.In ArtificialSocieties: The Com-
1996. Aligning simulation models: a case puterSimulationof SocialLife, ed. N Gilbert,
study and results. Comput. Math. Organ. R Conte, pp. 252-67. London:UCL Press
Theory.1(2):12341 Coveney P, Highfield R. 1995. Frontiers of
BainbridgeWS. 1995. Neuralnetworkmodels Complexity.New York:FawcettColumbine
of religious belief. Sociol. Perspect. 38(4): de Vos H, SmaniottoR, Elsas DA. 2001. Re-
483-96 ciprocalaltruismunderconditionsof partner
BainbridgeWS, Brent E, Carley K, Heise D, selection. RationalitySoc. 13(2):139-83
Macy M, et al. 1994. Artificialsocial intelli- DurkheimE. 1982 [1901]. TheRules of the So-
gence. Annu.Rev.Sociol. 20:407-36 ciological Method.London:Macmillan.
Bullnheimer B, Dawid H, Zeller R. 1998. Epstein J, Axtell R. 1996. Growing Artificial
Learningfrom own and foreign experience: Societies: Social Science from the Bottom
technological adaptationby imitatingfirms. Up. Cambridge,MA: MIT Press
Comput.Math. Organ. Theory4(3):267-82 Eshel I, HerreinerDK, SamuelsonL, Sansone
Caldwell SB. 1997. Dynamic Microsimulation E, Shaked A. 2000. Cooperation,mimesis,
and the Corsim 3.0 Model. Ithaca, NY: and local interaction.Sociol. Methods Res.
StrategicForecasting 28(3):341-64
CarleyK. 1991.A theoryof groupstability.Am. Flache A, HegselmannR. 1999. Rationalityvs.
Sociol. Rev. 56:331-54 learning in the evolution of solidarity net-
CartwrightD, HararyF. 1956. Structuralbal- works: a theoretical comparison. Comput.
ance: a generalization of Heider's theory. Math. Organ.Theory5(2):97-127
Psychol. Rev. 63:277-93 Flache A, HegselmannR. 2001. Do irregular
CastelfranchiC, Conte R, Paolucci M. 1998. grids make a difference?Relaxing the spa-
Normativereputationand the costs of com- tial regularityassumption in cellular mod-
pliance. J. ArtificialSoc. Soc. Simulat. 1(3): els of social dynamics.J. ArtificialSoc. Soc.
3.1-3.19 Simulat.4(4):6.1-6.27
CedermanL-E. 2001. Agent-basedmodelingin Flache A, Macy M. 1996. The weakness of
the political science. Polit. Methodol. 10(1): strongties: collective actionfailurein highly
xxx-xx cohesive groups.J. Math. Sociol. 21:3-28
ChattoeE. 1998. Justhow (un)realisticareevo- ForresterJW. 1971. WorldDynamics. Cam-
lutionaryalgorithmsas representationsof so- bridge,MA: MIT Press
cial processes?J. ArtificialSoc. Soc. Simulat. GilbertN. 1997. A simulationof the structure
1(3):2.1-2.36 of academic science. Sociol. Res. Online
Chwe MS-Y. 1999. Structureand strategy in 2(2):3.1-3.21
collective action.Am.J. Sociol. 105(1):128- Gilbert N, Troitzsch K. 1999. Simulationfor
56 the Social Scientist.Buckingham,UK: Open
Cohen MD, Riolo RL, Axelrod R. 2001. The Univ. Press.
role of social structurein the maintenanceof Halpin B. 1999. Simulationin sociology. Am.
cooperativeregimes. RationalitySoc. 13(1): Behav. Sci. 42(10):1488-508
5-32 Hanneman RA, Collins R, Mordt G. 1995.
 FROMFACTORSTO ACTORS 165

Discovering theory dynamics by computer LomiA, LarsenER.1998. Density delay andor-

simulation:experimentson state legitimacy ganizationalsurvival:computationalmodels
andimperialistcapitalism.Sociol. Methodol. and empirical comparisons.Comput.Math.
25:1-46 Organ.Theory3(4):219-47
HegselmannR, FlacheA. 1998. Understanding Macy M. 1991. Chains of cooperation:thresh-
complex social dynamics:a plea for cellular old effects in collective action. Am. Sociol.
automatabased modelling. J. ArtificialSoc. Rev. 56:73047
Soc. Simulat. 1(3):1.1-1.36 Macy M, Skvoretz J. 1998. The evolution of
Holland J. 1995. Hidden Order:How Adapta- trust and cooperation between strangers:a
tion Builds Complexity. Reading, MA: computationalmodel. Am. Sociol. Rev. 63:
Perseus 638-60
Homans GC. 1974. Social Behavior: Its Ele- MacLeodJ. 1995. Ain't No Makin'It. Boulder,
mentaryForms.New York:Harcourt,Brace, CO: Westview
Jovanovich MarkN. 1998. Beyond individualdifferences:
HopfieldJJ. 1982. Neuralnetworksand physi- social differentiationfrom first principles.
cal systemswith emergentcollective compu- Am. Sociol. Rev. 63:309-30
tationalabilities. Proc. Natl. Acad. Sci. USA Marwell G, Oliver P. 1993. The Critical Mass
79:2554-58 in CollectiveAction:A Micro-SocialTheory.
JohnsonPE. 1999. Simulationmodeling in po- Cambridge,UK: CambridgeUniv. Press
litical science. Am. Behav. Sci. 42:1509- McAdam D. 1988. Freedom Summer. New
30 York:OxfordUniv. Press.
Kaufman S. 1996. At Home in the Universe: MeadowsDL, BehrensWW m, MeadowsDH,
TheSearchfortheLaws of Self-Organization Naill RF,RandersJ, ZahnEK. 1974. TheDy-
and Complexity.Oxford, UK: Oxford Univ. namics of Growth in a Finite World.Cam-
Press bridge,MA: MIT Press
Kim H, BearmanP. 1997. The structureanddy- MeekerB, Leik R. 1997. Use of computersim-
namics of movement participation.Am. So- ulation for theory development: an evolv-
ciol. Rev.62:70-93 ing componentof sociological researchpro-
Kitts J, Macy M, Flache A. 1999. Structural grams. In Status, Network, and Structure:
learning:attractionand conformity in task- TheoryDevelopmentin GroupProcesses, ed.
orientedgroups.Comput.Math. Organ.The- J Szmatka,J Skvoretz,J Berger,pp. 47-70.
ory 5(2):129-45 Stanford,CA: StanfordUniv. Press
Klos TB. 1999. Decentralizedinteractionand Nowak A, SigmundK. 1998. Evolution of in-
co-adaptation in the repeated Prisoner's direct reciprocityby image scoring. Nature
Dilemma. Comput. Math. Organ. Theory 393:573-77
5(2):147-65 Nowak A, VallacherRR. 1998. Towardcompu-
Kollock P. 1993. An eye for an eye leaves tationalsocial psychology:cellularautomata
everyoneblind. Am. Sociol. Rev. 58:768-86 and neuralnetworkmodels of interpersonal
LataneB. 1996. Dynamic social impact:robust dynamics. In Connectionist Models of So-
predictionsfrom simple theory.In Modeling cial Reasoning and Social Behavior, ed. SJ
and Simulationin the Social Sciences from Read,LC Miller,pp. 277-311. Mahwah,NJ:
a Philosophy of Science Point of View,ed. R LawrenceErlbaum
Hegselmann,U Mueller, KG Troitzsch,pp. OrbellJ, Zeng L, MulfordM. 1996. Individual
287-310. Boston: KluwerDorderecht experience and the fragmentationof soci-
LomborgB. 1996. Nucleus and shield:the evo- eties. Am. Sociol. Rev. 61:1018-32
lution of social structurein the iteratedPris- PedoneR, ParisiD. 1997. In whatkindof social
oner's Dilemma. Am. Sociol. Rev. 61:278- groups can 'altruistic'behaviors evolve? In
307 SimulatingSocial Phenomena,ed. R Conte,
166 MACY ? WILLER

R Hegselmann,P Tera, pp. 195-201. Berlin: Schelling T. 1971. Dynamic models of segre-
Springer-Verlag gation.J. Math. Sociol. 1:143-86
Resnick M. 1997. TermitesTurtlesand Traffic SchumpeterJA. 1909. On the conceptof social
Jams. Cambridge,MA: MIT Press value. Q. J. Econ. 23:213-32
Reynolds CW. 1987. Focks, herds, and ShibutaniT. 1978. Derelicts of CompanyK:
schools: a distributed behavioral model. A Sociological Study of Demoralization.
Comput.Graphics21:25-34 Berkeley:Univ. Calif. Press
RosenkopfL, AbrahamsonE. 1999. Modeling Simon H. 1998. The Sciences of the Artificial.
reputationaland informationalinfluences in Cambridge,MA: MIT Press
thresholdmodels of bandwagoninnovation Smith TS, Stevens GT. 1999. The architecture
diffusion. Comput. Math. Organ. Theory of smallnetworks:stronginteractionanddy-
5(4):361-84 namic organizationin small social systems.
RummelhartD, McClelland J. 1988. Parallel Am. Sociol. Rev. 64:403-20
DistributedProcessing: Explorationsin the StrangD, Macy M. 2001. "In Searchof Excel-
Microstructure of Cognition. Cambridge, lence:" fads, success stories, and adaptive
MA: MIT Press emulation.Am.J. Sociol. 107(1):147-82
Saam N. 1999. Simulating the micro-macro TakahashiN. 2000. The emergenceof general-
link: new approaches to an old problem ized exchange. Am. J. Sociol. 105(4):1105-
and an applicationto militarycoups. Sociol. 34
Methodol.29:43-79 Vakas-DuongD, Reilley K. 1995. A system of
Saam NJ, HarrerA. 1999. Simulatingnorms, IAC neural networks as the basis for self-
social inequality, and functional change in organizationin a sociological dynamicalsys-
artificialsocieties. J. ArtificialSoc. Soc. Sim- tem simulation.Behav.Sci. 40:275-303
ulat. 2(1):2.1-2.20 WattsDJ. 1999. Networks, dynamics, and the
Sawyer K. 2001. Artificial societies and small-world phenomenon. Am. J. Sociol.
the micro-macro link in sociological the- 105(2):493-527
ory. Unpublished manuscript, Washington Willis P. 1977. Learning to Labor.New York:
Univ. ColumbiaUniv. Press