extensible multi-entity models a framework for incremental model construction

《extensible multi-entity models a framework for incremental model construction》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、ExtensibleMulti-EntityModels:AFrameworkforIncrementalModelConstructionKathrynBlackmondLaskeySuzanneM.MahoneyDepartmentofSystemsEngineeringInformationExtractionandTransport,Inc.MS5A61730N.LynnStreet,Suite502GeorgeMasonUniversityArlington,VA22209Fairfax,VA22030suzanne@iet.comklask

2、ey@gmu.eduAbstractGraphicalmodelshavebecomecommonforrepresentingprobabilisticmodelsinstatisticsandartificialintelligence.ABayesiannetworkisagraphicalmodelwhichencodesaprobabilitymodelasadirectedgraphinwhichnodescorrespondtorandomvariables,togetherwithasetofconditionaldistributio

3、nsofnodesgiventheirparents.InmostcurrentapplicationsofBayesiannetworks,afixednetworkisspecifiedtoapplytoallprobleminstances.Inferenceconsistsofconditioningoncertainrandomvariables,calledevidencevariables,andinferringthedistributionofothers,calledtargetvariables.Inmorecomplexprob

4、lemsarisinginartificialintelligence,itisusefultousethebeliefnetworkformalismtorepresentuncertainrelationshipsamongvariablesinthedomain,butitisnotpossibletouseasingle,fixedbeliefnetworktoencompassallprobleminstances.Thisisbecausethenumberofentitiestobereasonedaboutandtheirrelatio

5、nshipstoeachothervariesfromprobleminstancetoprobleminstance.Thispaperdescribesaframeworkforrepresentingprobabilisticknowledgeasfragmentsofbeliefnetworksandamethodforconstructingsituation-specificbeliefnetworksforparticularprobleminstances.1.IntroductionThewidespreaddissemination

6、ofthetechnologyofgraphicalmodelshasenabledaquantumleapinthecomplexityofproblemsthatcanbeaddressedbyBayesianmethods.Agraphicalmodelisaparsimoniousandmodularparameterizationforajointprobabilitydistributiononasetofrandomvariables.Theprobabilitydistributionisexpressedasaproductoffac

7、tors,whereeachfactorinvolvesonlyasmallnumberofvariables.Thefactorizationpermitsthedistributiontobespecifiedusingamanageablenumberofparameters.Computationallyefficientgeneralpurposemethodsexistforinferringorapproximatingconditionaldistributionsofsomevariablesinagraphicalmodelgive

8、nothervariables.Alargeliteraturehasgrownontheto