Agent-Based Computational Economics

Agent-BasedComputationalEconomics∗LEIGHTESFATSIONDepartmentofEconomics,IowaStateUniversity,Ames,[e-mail]perNo.1,RevisedAugust24,2003Abstract:Agent-basedcomputationaleconomics(ACE)isthecomputationalstudyofeconomiesmodeledasevolvingsystemsofautonomousinteractingagents.Startingfrominitialconditions,specifiedbythemodeler,thecomputationaleconomyevolvesovertimeasitsconstituentagentsre-peatedlyinteractwitheachotherandlearnfromtheseinteractions.ACEisthereforeabottom-upculture-dishapproachtothestudyofeconomicsystems.Thischapterdiscussesthekeycharac-teristicsandgoalsoftheACEmethodology.Eightcurrentlyactiveresearchareasarehighlightedforconcreteillustration.PotentialadvantagesanddisadvantagesoftheACEmethodologyareconsidered,alongwithopenquestionsandpossibledirectionsforfutureresearch.Keywords:Agent-basedcomputationaleconomics;autonomousagents;interactionnetworks;learning;evolution;mechanismdesign;computationalexperiments;object-orientedprogramming.1IntroductionDecentralizedmarketeconomiesconsistoflargenumbersofeconomicagentsinvolvedindistributedlocalinteractions.Theselocalinteractionsgiverisetomacroeconomicregularitiessuchastradingprotocols,sociallyacceptedmonies,andwidelyadoptedtechnologicalinnovationswhichinturnfeedbackintothedeterminationoflocalinteractions.Theresultisacomplicateddynamicsystemofrecurrentcausalchainsconnectingagentbehaviors,interactionnetworks,andsocialwelfareoutcomes.Thistwo-wayfeedbackbetweenmicrostructureandmacrostructurehasbeenrecognizedwithineconomicsforaverylongtime[42,75,88,91].Formuchofthistime,however,economistshavelackedthemeanstomodelthisfeedbackquantitativelyinanywherenearitsactualcomplexity.Themostsalientcharacteristicoftraditionalquantitativeeconomicmodelssupportedbymicrofounda-tionsistheirtop-downconstruction.Heavyrelianceisplacedonexternallyimposedcoordinationdevicessuchasfixeddecisionrules,commonknowledgeassumptions,representativeagents,andmarketequilibriumconstraints.Face-to-faceinteractionsamongeconomicagentstypicallyplaynoroleorappearintheformofhighlystylizedgameinteractions.Inshort,economicagentsinthesemodelshavelittleroomtobreathe.∗Thissurveyisamodifiedversionof[103],whichwasdirectedmorespecificallytoartificialliferesearchers.1Inrecentyears,however,substantialadvancesinmodelingtoolshavegreatlyexpandedthepossibilitysetforeconomists[8,14,27,34,43,58,87,111].Researcherscannowquantitativelymodelawidevarietyofcomplexphenomenaassociatedwithdecentralizedmarketeconomies,suchasinductivelearning,imperfectcompetition,endogenoustradenetworkformation,andtheopen-endedco-evolutionofindividualbehaviorsandeconomicinstitutions.Onebranchofthisworkhascometobeknownasagent-basedcomputationaleconomics(ACE),thecomputationalstudyofeconomiesmodeledasevolvingsystemsofautonomousinteractingagents.1ACEresearchersrelyoncomputationalframeworkstostudytheevolutionofdecentralizedmarketeconomiesundercontrolledexperimentalconditions.TwobasicconcernsdriveACEresearch.Oneconcernisdescriptive,focusingontheconstructiveexplanationofemergentglobalbehavior.Whyhaveparticularglobalregularitiesevolvedandpersistedinreal-worlddecentralizedmarketeconomies,despitetheabsenceoftop-downplanningandcontrol?How,specifically,havetheseglobalregularitiesbeengeneratedfromthebottomup,throughtherepeatedlocalinteractionsofautonomousinteractingagents?Andwhytheseparticularregularitiesandnotothers?Thesecondconcernisnormative,focusingonmechanismdesign.Givenaparticulareconomicmechanism,whetherexistingorsimplyenvisioned,whataretheimplicationsofthismechanismfortheperformanceoftheeconomyasawhole?Whatsocialoutcomeswillresultfromtherepeatedattemptsofself-seekingagentstoexploitthemechanismtotheirownadvantage?Asinaculture-dishlaboratoryexperiment,theACEmodelerstartsbyconstructinganecon-omycomprisinganinitialpopulationofagents.Theseagentscanincludebotheconomicagents(e.g.,consumers,producers,intermediaries,...)andagentsrepresentingvariousothersocialandenvironmentalphenomena(e.g.,governmentagencies,landareas,weather,...).TheACEmodelerspecifiestheinitialstateoftheeconomybyspecifyingtheinitialattributesoftheagents.Theini-tialattributesofanyoneagentmightincludetypecharacteristics,internalizedbehavioralnorms,internalmodesofbehavior(includingmodesofcommunicationandlearning),andinternallystoredinformationaboutitselfandotheragents.Theeconomythenevolvesovertimewithoutfurtherinterventionfromthemodeler.Alleventsthatsubsequentlyoccurmustarisefromthehistoricaltime-lineofagent-agentinteractions.Noextraneouscoordinationdevicesarepermitted.Forex-ample,noresortcanbemadetotheoff-linedeterminationandimpositionofmarket-clearingpricesthroughfixedpointcalculations.ThefollowingsectionsprovideamoredetaileddiscussionoftheACEmethodology.Beforeproceeding,however,animportantdisclaimerisinorder.Theprimaryobjectiveofthissurveyismodestinscope:tointroduce,motivate,andillustratethroughconcreteexamplesthepotentialusefulnessoftheACEmethodologybyhighlightingselectedpublicationsineightcurrentlyactiveresearchareas.Thenu