资源描述:
《running head learning nested agent models.》由会员上传分享,免费在线阅读,更多相关内容在教育资源-天天文库。
1、LearningNestedAgentModelsinanInformationEconomyJoseM.VidalandEdmundH.DurfeeArticialIntelligenceLaboratoryUniversityofMichigan1101BealAvenueAnnArbor,MI48109-2110313-763-9074,313-936-1563fjmvidal,durfeeg@umich.eduMay26,1997RunningHead:LearningNestedAgentModels.AbstractWepresentourappro
2、achtotheproblemofhowanagent,withinaneconomicMulti-AgentSystem,candeterminewhenitshouldbehavestrategically(i.e.learnandusemodelsofotheragents),andwhenitshouldactasasimpleprice-taker.Weprovideaframeworkfortheincrementalimplementationofmodelingcapabilitiesinagents,andadescriptionoftheform
3、sofknowledgerequired.Theagentswereimplementedanddierentpopulationssimulatedinordertolearnmoreabouttheirbehaviorandthemeritsofusingandlearningagentmodels.Ourresultsshow,amongotherlessons,howsavvybuyerscanavoidbeingcheated"bysellers,howpricevolatilitycanbeusedtoquantitativelypredictthe
4、benetsofdeepermodels,andhowspecictypesofagentpopulationsin
uencesystembehavior.1IntroductionInopen,multi-agentsystems,agentscancomeandgowithoutanycentralcontrolorguidance,andthushowandwhichagentsinteractwitheachotherwillchangedynamically.Agentsmighttrytomanipulatetheinteractionstothe
5、irindivid-ualbenets,atthecostoftheglobaleciency.Toavoidthis,theprotocolsandmechanismsthattheagentsengageinmightbeconstructedtomakemanipulationirrational(RosenscheinandZlotkin,1994),butunfortunatelythisstrategyisonlyapplicableinrestricteddomains.Bysituatingagentsinaneconomicsociety,as
6、wedointheUniversityofMichiganDigitalLibrary(UMDL),wecanmakeeachagentresponsibleformakingitsowndecisionsaboutwhentobuy/sellandwhotodobusinesswith(Atkinsetal.,1996).Amarket-basedinfrastructure,builtaroundcomputationalauctionagents,servestodiscourageagentsfromengaginginstrate-gicreasoning
7、tomanipulatethesystembykeepingthecompetitivepressureshigh.However,sincemanyinstancescanarisewhereimperfectionsincompetitioncouldbeexploited,agentsmightbenetfromstrategicreasoning,eitherbymanipulat-ingthesystemornotallowingotherstomanipulatethem.Butstrategicreasoningrequireseort.Ana