The events of September 11,2001, were a wake-up call forAmericans. All of us remember

where we were and what we weredoing when we learned of the horrificattacks. On that terrible day, I waspresenting results of a financial agent-based simulation to corporate execu-tives on the top floor of a Houstonskyscraper. I had no idea why theUnited States was attacked, but I hadplenty of time to contemplate thequestion as I drove back to Santa Fein my rental car. What had we done tomotivate terrorists to take such hatefulactions against us? What messagewere they trying to send us? Why didthese attacks happen?

As the world strives to understandterrorism better and struggles to iden-tify its many faces and forms, fewwould deny that religiously motivatedterrorism is becoming increasinglyprevalent. But the association of reli-gion with violent, radical groups,many of which have their own inter-pretation of a religion, needs to beexamined carefully. The violentactions of radical Islamist groups, forexample, have led to the mistaken

association of terrorism with Islam.One-fifth of the world population is

Islamic. Dispersed around the globe,the largest concentrations of Muslimsare in Indonesia, Pakistan, India,Bangladesh, and the Middle East. InWestern society, the largest concentra-tion is in the United States. The reli-gion itself is based upon peace (theword Islam means “self-surrender” inArabic, and the universal greeting ofMuslims is “salaam alaikum,” whichmeans “peace be upon you”). In manyIslamic societies, however, the passiveor neutral behaviors of the peacefulmajority often become obscured by theattention-seeking acts of a “noisyminority.” Although the point isdebated, the general understanding isthat most Muslims are peacefulbecause of their Islamic beliefs, andthat the “noisy minority” has misinter-preted Islamic teachings.

Muslims are taught thatMuhammad was sent by Allah tospread belief in a single God—asopposed to the multitude of pagan rit-uals honoring a variety of deities at histime (A.D. 610). For many modern-day Muslim radicals, especially those

in traditional societies, American popculture may be perceived as beingsimilar to old-fashioned paganism, acult that worships money and sex.Some modern-day militants may per-ceive themselves as following a pathsimilar to Muhammad’s in cleansingthe Islamic world from the infiltrationof the “pagan” West. Other Islamicpeople may fear that their culture, tra-ditions, and beliefs are being replacedbecause globalization imposesWestern values on them. Being mutu-ally understanding of religious sensi-tivities, as well as responsible andrespectful of each other’s influence,will help establish a peaceful coexis-tence between the West and theIslamic world. Finding ways to movetoward this goal requires carefulanalysis and discussion.

The Complex Systems Group atLos Alamos has been examining ques-tions related to the “why” behind ter-rorist organizations in the MiddleEast. Borrowing tools from the fieldof computational economics and soci-ology, we are developing agent-basedmodels that simulate social networksand the spread of social grievances

184 Los Alamos Science Number 28 2003

Understanding Why–Dissecting radical Islamist terrorism

with agent-based simulation

Edward P. MacKerrow

within those networks. Our computer-generated “agents” are humanlike,endowed with personal attributes andallegiances that statistically match thedemographics of a specified regionand, like people, interact with oneanother and respond to societal pres-sures. The defining feature of ouragents, however, is that their “behav-iors” are allowed to change during asimulation run. For example, an agentmay “learn” during the simulation notto interact with agents of a certainsocial class, or an agent may developdeep “feelings” of oppression andgrievance based on its experiences.

We do not know a priori the lifestories of our agents, but after tens ofthousands have interacted, we haveproduced a scenario, or a virtual his-tory, for a region of interest. The plau-sibility of this scenario is normallyassessed by human experts who havecomplete knowledge of the modelassumptions and the rules followed bythe agents. By replaying any particu-lar simulation, the experts can observehow the agents behaved and examinewhy they behaved in a certain way.

We can expose our agents to a vari-ety of determinants—new governmentpolicies, different media exposure,economic pressures, and others—andquickly generate hundreds of new sce-narios. Thus, we can conduct compu-tational experiments that can beanalyzed statistically and objectivelyto increase our insight, support deci-sion making, and aid policymakers(see Figure 1). Scenarios can even beused to gain insight into actual eventsthat have little or no historical prece-dence. It should be emphasized, how-ever, that the goal of these simulationsis not to predict specific events andnot to estimate the probability or fre-quency of terrorist acts, but to gener-ate scenarios and analyze them.

Our work is part of the DefenseThreat Reduction Agency (DTRA)research effort known as the ThreatAnticipation Program (TAP).1 Built

on a multidisciplinary team of schol-ars from the arts and literature, historyand psychology, Middle Eastern andMuslim cultures, religion, economics,and sociology, TAP aims to developalgorithms and software frameworksthat can generate the most likely mod-els of terrorism and terrorist scenariosin order to catch the precursor signalsof the next terrorist attack. We hopethat TAP can eventually provide uswith insight into potential prevention,interdiction, and mitigation policies.

Modeling ComplexSocioeconomic Systems

The social tensions in the MiddleEast emanate from many different yetinterrelated conflicts, and each MiddleEastern nation has a unique history inrelation to those conflicts (Miller1997). Therefore, the underlyingsocial processes cannot be understoodby a simple linear combination of sep-arate sociologic, economic, demo-graphic, religious, cultural, andpolitical subprocesses.

Agent-based simulation providesus with a methodology for modelingcomplex socioeconomic phenomena.Agent-based simulation was first

Number 28 2003 Los Alamos Science 185

Understanding Why

1TAP was originated by Dr. StephenYounger, director of DTRA. Younger is aformer director of the nuclear weaponsprogram at Los Alamos.

Media

Adrar1

Alger

Matruh

Oran

Al Qahirah (Cairo)

Constantine

Tendouf

0–0.03 0.14–0.17 0.31–0.35

Perceived hardship

Num

ber

of a

gent

s w

ith p

erce

ived

har

dshi

p le

vel

0.49–0.52 0.66–0.70

16

14

12

10

8

6

4

2

0

Figure 1. A Sample Distribution from the TAP ModelThe TAP model places thousands of agents throughout the Middle East, endowsthem with numerous properties and behaviors, and allows them to interact for num-ber of simulation years, thus creating a short, virtual history of the region. We candissect that history and analyze various social metrics (such as perceived socialdisadvantage, directed grievances, and allegiances) to gain insight into extremistbehaviors. The figure shows a distribution for perceived hardship, a metric thatmight be obtained in reality by a pollster asking the question, “On a scale of 0 to 1,what is your level of hardship?”The colors correspond to some districts in Algeriaand Egypt. Because agent behaviors are not preprogrammed, each simulation usingthe TAP model will produce a different distribution of perceived hardship (or anyother social metric). Human experts must assess whether the envelope of results isvalid, that is, whether our model correctly brackets the possible levels of real-worldsocial grievances. That task is difficult because the model can be validated withonly one data point—the real-world poll made under specific conditions.

introduced into economics to addressshortcomings with economic simula-tions, which in early versions assumedhomogeneous populations of ideal-ized, perfectly rational agents whohad perfect information about perfectmarkets. The results of those simula-tions, though frequently used, wereoften incorrect because of the flawedrepresentation of real-world agentbehavior. See Shubik (1997) for a dis-cussion of issues associated withgame theory applied to real-worldapplications.

As the computer became further inte-grated into the social sciences, morerealistic socioeconomic models wereattempted, and the methods of agent-based mathematics began to develop.Agent-based models have now foundwidespread use in economics and allowagents to act with bounded rationality,based on imperfect or incomplete infor-mation, and to act on chance and per-ceived economic utility. In addition tomore realistic representations of individ-uals, agent-based simulation allows foranalysis of nonequilibrium conditionscompared with the historical practice ofanalysis made at equilibrium points—many real-world socioeconomic systemsare not in equilibrium and may neverreach equilibrium.

Currently, the major difficulty weface in building a model of a com-plex socioeconomic system is inquantifying social situations. Notonly do we need better models thatshow how to represent social interac-tions, but we also need better empiri-cal analysis of actual real-worldstudies. A fundamental problem isthat real-world “observables” may begenerated by many different interac-tion processes; therefore, empiricalfindings are open to different inter-pretations. My belief is that certainsocial micromodels apply better thanothers, depending on the context inwhich they are applied—and a goodmodel will utilize a broad suite ofsocial micromodels.

The Los Alamos TAP Agent-Based Model

The TAP simulation is built withobject-oriented software and imple-mented in the Java programming lan-guage. (The development ofobject-oriented software has beendirectly related to rapid advances inagent-based simulation over the last15 years.) Although, in theory, any

computer language can be used torepresent a social system, object pro-gramming fits naturally with model-ing social systems and greatly reducesdevelopment time.

In a simplified description, we firstestablish the initial properties of“objects.” The most important objectsin the TAP model are Nation,District, Organization,Mosque, and Person. (In this arti-

186 Los Alamos Science Number 28 2003

Understanding Why

Objects

States

Member actions

CellLeaderRole

void callMeetingvoid listenForInstructions

PersonRole

Person meetRandomPersonPerson meetSocNetPerson

Person

int ageboolean maleint educationLevelString religionString ethnicityfloat religiousExtremismint pedigreeLevelint incomeLevelfloat perceivedHardshipPersonRole[ ] rolesfloat socialGrievancefloat riskAversionLevel

void updateSocialGrievancevoid updatedPerceivedHardship

...

...

WeaponsEngineerRole

void assemblePortfolio

Portfolio weaponsPortfolioint experienceLevel

RecruiterRole

void solicitPerson recruit

StudentRole

void attendSchool

String schoolName

TerroristRole

void collectResourcesvoid spendResourcesvoid contactLeader

TerroristOrganization

TerroristCell getCellvoid sendMessage

InterfaceOrganizationListener

void listenForInstructions

Organization

Position headquartersPerson leaderAllegianceVector vectorArrayList cells

TerroristCell

HashMap membersPosition locationPersonRole leader

Organization getParentOrg

Figure 2. Objects in the TAP ModelObject-oriented programming is a natural fit to agent-based modeling because wecan design objects that learn and adapt based on their history, their current state,and the states of other objects. This class diagram shows some of the object typesused in the simulation. The simulation is built upon many different instances ofthese object types, each with different attributes. The object architecture allows forflexibility; the PersonRole class, and its inherited subclasses, allow a constructwhere any one Person object can play multiple roles. Interfaces allow for specifica-tion of required actions that can be implemented differently, depending upon thetype of object implementing the interface (interface relationships are shown byblack dotted lines. Objects can be composed of other objects allowing for newobjects to build upon the structure of existing objects (shown by blue dotted lines).

cle, we identify all software objectsby capitalization and by using aCourier font.) Each object containsattributes that support a correspondingabstraction representation and modesof actions that represent agent behav-iors (see Figure 2). We divide eachMiddle Eastern nation into a series ofadministrative District objects, asdefined by standardized GeographicalInformation System data (seeFigure 3). Each District is thenpopulated with a number of agents.Empirical distributions derived from

regional demographic and ethno-graphic data (see Table I ) are used toinitialize agent attributes. Differenttypes of agents are instantiated withdifferent data sets and different rulesof behavior.

For each District, we definerelative weighting factors to estimatethe “social welfare,” or “social capi-tal,” of religious and ethnic groups inthat region. We posit that social capi-tal is a weighted sum of income, eth-nicity, religion, education, andpedigree, where pedigree represents

inherited or appointed social wealth.(A Saudi prince would have a veryhigh pedigree value.) We allow forflexibility of social weighting factorsacross different districts; for example,in some districts, ethnicity may notinfluence social status as much asdoes income.

We also take into account theimportant aspect of social rank in asociety. Each agent updates bothsocial capital and social rank on a reg-ular basis during a simulation run. Asdiscussed later in the text, social rankplays an important role in the theorywe use to model social interactions—although by itself it is not necessarilya determining attribute of a terrorist.

Modeling Interactions

Many of the behaviors and opin-ions of individuals are rooted in thesocial structures to which theybelong. For example, a young adult’sproclaimed dislike toward the UnitedStates may be socially inherited fromparents and reinforced through fam-ily circles, friendship circles, and bythe media. Understanding whichsocial experiences, conditions, andinteractions increase the likelihood ofbecoming a terrorist is very difficult.Lay explanations for an individual’schoice to become a terrorist includedesperation, poverty, mental illness,and lack of education. Statistics ofsuicide attackers, however, show thatlow income levels are neither neces-sary nor sufficient to explain suicideattacks. Education is not a determin-ing factor either, and ironically somedata may suggest that higher educa-tion levels are positively correlatedwith terrorist attributes. So what is itthen? The answer may be found by acloser examination of the terroristorganizations themselves.

Many terrorist organizations usecharismatic leaders to cultivate andindoctrinate small cells of young

Number 28 2003 Los Alamos Science 187

Understanding Why

Figure 3. Region and Other Objects in the TAP ModelA screen capture from the TAP model shows many of the geographical regionobjects used to delineate different groups of agents as well as other objects. Eachregion is described by a different demographic composition. Agents are instanti-ated inside each region on the basis of these regional demographics. The whiteareas in northern Egypt, for example, indicate clusters of agents.

Table I. Attributes Used to Set the State of Agents

Agent Attribute Distribution Type Data Source

Age Empirical (discrete) U.S. Census BureauSex Empirical (male, female) U.S. Census BureauEducation Empirical (years) CIA World FactbookEducation type Estimated (religious, secular) CIA World FactbookEthnicity Empirical (percent by group) CIA World Factbook Religion Empirical (percent by group) CIA World FactbookExtremism Estimated ([0,1]) Interviews and readingsPedigree Estimated ([0,1]) Interviews and readingsIncome Empirical ($ per year) World BankMarried Empirical (Boolean) The Economist Employment Empirical (Boolean) The Economist, World Bank Location Empirical (# per km2) GIS Data Sources

recruits to become martyrs for theoverall cause (Atran 2003). (Note that,for terrorist organizations, the use ofmartyrs is very economical, approxi-mately $150 per suicide bomberattack.) These leaders, who rarelybecome martyrs themselves, seek toestablish a commitment between themembers of the cells in the form of asocial contract—usually sealed with avideo testimony. A cell member devel-ops a sense of obligation to the fictive“kinship” of the cell. The social costto individuals for reneging on com-mitments is very high; they risk beinglabeled “kafir”—an infidel, or nonbe-liever. Defectors from some terroristorganizations can even be killed bythe organization. Peer pressure fromthe cell is a motivating factor forsome terrorists; others are fully com-mitted on their own and believestrongly in becoming terrorists. Thelatter appears to be particularly truewith respect to terrorists who opposeIsraeli settlements in the occupied ter-ritories of Palestine. In the TAPmodel, we include different types ofagents and allow some agents to bemore self-motivated toward terrorismand others to require more or lessactive recruitment.

Structural and institutional con-straints induce individuals to act in amanner most consistent with the pref-erences of the social structure or insti-tution—in this case, the terroristorganization. In the TAP model, weintroduce these constraints usingsocial network representations forestablishing probabilistic social rule-sets. To include realistic interactions,we construct several different types ofsocial networks between agents: kin-ship, religious, organizational, andfriendship. Although meetings canoccur between any two agents orbetween an agent and a group, weassume that interactions are moreprobable between agents on the samesocial network than between randomagents in the local population, and we

weight the interactions accordingly.A screen-capture image of an initial

friendship network of agents is shownin Figure 4(a), where we have used thealgorithm of Jin et al. (2001) to con-struct the network. It is obvious howdense the friendship network is at theend of the preprocessing stage—eachagent has a large number of socialcontacts. During the following stagesof the simulation, this network willevolve into cliques and a less uniformdensity as agents evolve their socialnetworks—as shown in Figure 4(b).

We include simulated social net-works as structures that affect socialinteractions in our model, althoughwe have also included the ability topopulate the model with known ter-rorist networks. Obtaining data andcharacteristics on the actual socialstructures of the various terroristorganizations around the world is dif-ficult because of their covert nature.In place of existing terrorist networks,we use surrogate information tomodel these social networks. (Anexample of a surrogate network isgiven in the box, “An al-QaidaNetwork” on the opposite page.)

Interacting Agents

When an agent meets another agentin our simulation, a social interchangeoccurs. Depending on the outcome ofthe meeting, a certain amount of inter-active learning occurs. Because inter-actions are part of the larger socialstructure, a form of “social learning”occurs throughout the agent popula-tion—observable in part through theirheterogeneous allegiances.

Each agent in the TAP simulationcarries an “allegiance vector.” The ele-ments of this vector contain an integervalue representing that agent’s alle-giance toward specific nations, organi-zations, ethnic groups, or religiousgroups. A positive (or negative) valueof allegiance suggests a positive (ornegative) allegiance for the group asso-ciated with that element. Figure 5(a)shows a schematic of the allegiancevector concept.

The Gallup Polls of the MiddleEast are aids to understanding the wayIslamic nations feel toward othernations on a variety of topics. We usethese polls in the TAP model for esti-mating our aggregate allegiance vec-

188 Los Alamos Science Number 28 2003

Understanding Why

Figure 4. Evolving Networks(a) This screen capture shows a large friendship network of agents in the TAP modelat the end of the preprocessing stage. At this point in time, the network is ratherdense; each agent is connected to an average of five other agents. A few agents ini-tialized as Armed Islamic Group (GIA) terrorists are shown in red. (b) As agentsinteract with other agents during the simulation, the network evolves. The colors ofthe lines connecting the agents reveal how the agents became friends—eitherthrough random meetings or through mutual friendships. Some agents become iso-lated in stranded cliques and associate with very few other agents.

(a) (b)

tors. Figure 5(b) shows an example ofdata from a Gallup poll.

Social Bargaining and the NashDemand Game. Many theories ofsocial interaction can be used for

modeling ways in which beliefs,opinions, and values are communi-cated, shared, and modified duringsocial meetings. We require a theorythat allows us to model how alle-giance values are transferred between

interacting agents.One theory we use is a social bar-

gaining theory, established byH. Peyton Young (1998), that is basedon the one-shot Nash demand game.During an interaction, each one oftwo agents places a bid for some por-tion of an abstract available “prop-erty,” where the bid is related to theestimated value of establishing asocial contact with the other agent.Both parties get their demands if thesum of the bids is less than the totalproperty available; that is, successfulbargaining occurs if the two agentsare not too greedy. If one agent’s bidis bi and the other’s is bj, then a suc-cessful bargaining process occurs if bi + bj ≤ 1. The condition for a pureNash equilibrium (where the agentsare at their best-bid positions, and achange in bid by either agent willlower the overall payoff) is bi + bj = 1.

Agents learn from past interactionshow to bid optimally in Nash demandgames. Each agent retains a memoryof its past m meetings. Agents associ-ate some attributes of those agentsthat they interacted with in these pastm meetings and judge how well theydid in demand games with agents ofsimilar attribute types. This informa-tion is then used as a basis for futurebids with agents of similar type—in aform of social learning.

If the one-shot social bargaininggame is successful, then allegiancevalues are transferred in an asymmet-ric, bidirectional manner between thetwo agents. The agent with the lowersocial status “absorbs” more of theother agent’s allegiance values,whereas the agent with the higher sta-tus absorbs fewer. In this way, socialnorms can emerge from the continuedinteractions between agents. Note thatthis is but one heuristic from a set of“social rules” used by the agents.

We hope that this methodologywill represent one component relatedto the spread of beliefs through socialstructures. Some Islamic organizations

Number 28 2003 Los Alamos Science 189

Understanding Why

An al-Qaida Network

The compartmentalized social network of the al-Qaida cells involved inthe September 11 hijackings, researched by Valdis Krebs (2002), isshown above. Krebs obtained open-source data on the hijackers as thosebecame available after the attacks. Although some nodes are likely miss-ing, his analysis and construction of the network are very useful in under-standing some features of terrorist cells.

The social network of the hijackers was very loosely connected andsparse. Whereas Mohamed Atta is clearly seen as the ringleader, many ofthe hijackers were separated by a few degrees—more than one step awayfrom each other. This separation even applied to hijackers on the sameflight. The strategy ensures that the entire network is robust to the captureor compromise of a cell member. Usama bin Laden described this strat-egy on a videotape that was found in a deserted al-Qaida house inAfghanistan. This type of covert social network suffers from reducedinformational efficiency and information sharing, although the hijackerswere clearly able to mitigate those deficiencies and ensure some level ofcommunication and resource planning. (Courtesy and permission of Valdis Krebs.)

establish relief efforts and supplyresources for citizens of low socioeco-nomic status. In addition to providingneeded welfare, these actions alsoimprove public support of and opinionabout Islamic organizations. Someorganizations rely upon a “bottom-up”approach to instill the populace withtheir doctrine (that is, Islamic Law—Shari’a), whereas others may choose amore revolutionary “top-down”approach by replacing a secularregime with a more Islamic rulingparty. The social bargaining approachfits closer with the “bottoms-up”approach of belief spreading.

Social Repression andEstimation of Social

Grievance

I have briefly outlined how alle-giances are transferred through agentpopulations by social learning. Thislearning is affected by who meetswhom, how socioeconomic status isvalued relative to allegiance adoptionthat occurs between agents, and pastexperiences of agents in meeting withagents that are considered similar. Atthe end of the simulation, however,we want to quantify relative measuresof propensity toward terrorism acrossthe agent population.

Social grievance is one of the“summary” metrics we use to deter-mine when an agent, or a collection ofagents in a region or organization, isconsidered to have a propensity forprotest and therefore has a higherpotential to become a terrorist. It iscalculated from metrics that includean agent’s sense of social repression.Social grievance is directed against aparticular group or organization. Inthe simulation, repression stems fromsocial disadvantage, inherited alle-giances, cultural penetration, repres-sion from the regime, and mediainfluences.

We calculate a composite socioeco-

nomic disadvantage metric based on theagent’s social ranking and the socialranking of the groups that the agentidentifies with most. An agent mayhave high social status but may identifywith a social group that has low socialstatus—thereby increasing the level ofdisadvantage felt for other members ofthe group.

Agents perceive repression by acorrupt regime. We quantify this levelof repression by calculating the over-

lap between the agent’s and theregime’s allegiance vectors, weightedby a corruption factor. Corrupt gov-erning regimes in the Middle East thatare secular and aligned with Westernnations for commercial trade and mili-tary support, and who do not supportthe underlying social welfare of theirpopulations, are considered to be animportant root cause of militantIslamist terrorism against the West.

The infiltration of some aspects of

190 Los Alamos Science Number 28 2003

Understanding Why

100

80

60

40

20

0

-20

-40

-60

-80

Alig

nmen

t

Egy

pt

Jord

an

Ham

as

PIJ

UK

US

A

Iran

Iraq

Liby

a

Hiz

balla

h

EIJ

al-A

qsa

al-Q

aida

Figure 5. Typical Allegiance Vector(a) All agents have allegiance vectors, which contain more than 100 elements. Thevalues of individual vector elements indicate the feelings of an agent with regard tocountries, organizations, and groups. These values change as agents interact witheach other, thus simulating the process of social memes or “contagion” effects.(b) The results of Gallup Middle Eastern opinion polls help us quantify allegiancevalues in our model.

United States

Great Britian

France

Russia

China

Veryfavorable

Somewhatfavorable

Neitherfavorable norunfavorable

Somewhatunfavorable

Veryunfavorable

Don't know/no answer

3 11% 14% 9% 54% 9%

4 14% 19% 14% 40% 9%

7 31% 24% 10% 17% 11%

3 26% 30% 11% 19% 11%

6% 35% 26% 10% 12% 11%

Western culture, whether physically ormedia based, into Islamic regions isconsidered to be another source ofrepression to Islamic groups. In theTAP model, we estimate the contribu-tion of cultural penetration to socialrepression as a weighted sum of thefraction of the local population that isnonindigenous and a media presencefactor, which represents the relativeamount of external (foreign) mediainfluence in that region. The weight-ing function depends on the relativeallegiance values—influences fromcultures with high allegiance valuesare considered good. We estimate themedia influence factor from “surro-gate” data. For many Middle Easternnations, we have measures of the per-centage of households that haveaccess to television, radio, and theInternet. Though not always the case,a household that has access to theseinformation sources usually hasaccess to the media of all other cul-tures. (In some Middle Easternregions, the regime may censor exter-nal media influences.)

After calculating an agent’s socialdisadvantage and level of repression,we calculate agent A’s time-dependentgrievance G toward the social groupM as

G = d(t) × a(M, t) × o(t) × p(t) × f(M),

where the first term is the socioeco-nomic disadvantage of A, the secondterm is the dislike A feels for groupM, the third and fourth terms are theregime and cultural-penetration con-tributions to A’s perceived oppression,and the last term is a measure of A’sperception of group M’s level of cor-ruption. Social grievance of the agentsin each region is monitored in order toobtain a social grievance potential asthe simulation progresses. An indica-tor of terrorist instability exists ifagents with high levels of socialgrievance have access (through theirsocial networks) to a terrorist organi-

zation with similar grievance targets.This represents increased public sup-port and more probable recruitmentfor terrorist organizations.

Summary

I have briefly described a complexagent-based model of a complex situa-tion—terrorism associated with mili-tant groups. My hope is to give aflavor of the methodology used in con-structing agent-based simulations ofsocioeconomic systems and to showhow this methodology is being appliedto the challenges in developing adetailed understanding of the sociody-namics of militant Islamist terrorism.

I see agent-based simulations ascomputational experiments that con-vey a great deal of scenario informa-tion in a timely, efficient, and safemanner. Examining the path-depend-ent time evolution of a particularlyinteresting simulation result (a virtualhistory) in replay mode allows us toanalyze the “how” and “why” of agentbehavior. If the TAP agent-basedmodel can supplement policymakingin the turbulent clash between theWest and Islamist radicals and canhelp policy makers visualize andunderstand important yet currentlyunknown interrelationships, then thiswork will be a success. Althoughagent-based simulation can help usgain insight into complex systembehavior, my longer-term hope is thatthe “tit-for-tat” pattern of violence inthe Middle East will be quenched bytaking an honest look at both sides ofthe issues at hand through peacefulnegotiation and mutual respect.

Acknowledgments

I thank the DTRA AdvancedSystems and Concepts Office for theirkind support and valuable expertise inconstructing the TAP model. I also

thank Gary Ackerman, AnjaliBhattacharjee, and many others at theMonterey Institute of InternationalStudies for sharing their valuableexpertise on terrorism and the MiddleEast; Charles Macal and MichaelNorth of Argonne NationalLaboratory for their expert modelingadvice; and the Santa Fe Institute forits support and useful input; AmirMohagheghi, Faraj Ghanbari, JessicaTurnley, and Nancy Kay Hayden ofSandia National Laboratories for theiruseful discussions. I thank WaleedEl-Ansary of the Islamic ResearchInstitute for his useful insights intoIslamic economics and game-theory.Finally, I thank the Los Alamos teamof Joe Holland, Karl Lautenschlager,Merle Lefkoff, Ayla Matanock,Dennis Powell, Brian Reardon, DavidSharp, and Zoltan Toroczkai for theirimportant contributions to this work. �

Number 28 2003 Los Alamos Science 191

Understanding Why

Edward MacKerrow graduated from San JoseState University in 1986 with a bachelor’sdegree in physics, from the University of NewMexico in 1989 witha master’s degree inphysics, followed in1995 by a Ph.D. inphysics. In 1989, Edcame to theLaboratory as aresearch graduateassistant. He contin-ued working at LosAlamos while doinghis Ph.D., thenbecame a postdoctoral fellow, and later a staffmember. In 2000, Ed took a leave of absence towork for BiosGroup, Inc., where he specializedin building agent-based models of both U.S.and international corporations in the banking,finance, and energy sectors, with an emphasison operational risk issues. He returned to LosAlamos in 2002 to continue his research inagent-based socioeconomic and organizationalmodeling and simulation, as a technical staffmember in the Theoretical Division’s ComplexSystems Group.