资源描述:
《Graphical Models》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库。
1、GraphicalModelsJinlongWu&TiejunLiNov20081IntroductionJordan[3]presentesaveryconciseintroductiontoGraphicalModels(GMs):Graphicalmodelsareamarriagebetweenprobabilitytheoryandgraphtheory.Theyprovideanaturaltoolfordealingwithtwoproblemsthatoccurthroughoutapplied
2、mathematicsandengineering—uncer-taintyandcomplexity—andinparticulartheyareplayinganincreasinglyimportantroleinthedesignandanalysisofmachinelearningalgorithms.Fundamentaltotheideaofagraphicalmodelisthenotionofmodularity—acomplexsystemisbuiltbycombiningsimpler
3、parts.Probabilitytheoryprovidesthegluewherebythepartsarecombined,ensuringthatthesystemasawholeisconsistent,andprovidingwaystointerfacemodelstodata.Thegraphtheoreticsideofgraphicalmodelsprovidesbothanintuitivelyappeal-inginterfacebywhichhumanscanmodelhighly-i
4、nteractingsetsofvariablesaswellasadatastructurethatlendsitselfnaturallytothedesignofefficientgeneral-purposealgorithms.Manyoftheclassicalmultivariateprobabilisticsystemsstudiedinfieldssuchasstatistics,systemsengineering,informationtheory,patternrecognitionands
5、tatisticalmechanicsarespecialcasesofthegeneralgraphicalmodelformalism—examplesincludemixturemodels,factoranalysis,hiddenMarkovmodels,KalmanfiltersandIsingmodels.Thegraphicalmodelframeworkprovidesawaytoviewallofthesesystemsasinstancesofacommonunderlyingformali
6、sm.Thisviewhasmanyadvantages—inparticular,specializedtechniquesthathavebeendevelopedinonefieldcanbetransferredbetweenresearchcommunitiesandexploitedmorewidely.Moreover,thegraphicalmodelformalismprovidesanaturalframeworkforthedesignofnewsystems.GMsareusuallydi
7、videdintotwotypes—undirectedanddirected.UndirectedGMsarealsocalledMarkovNetworksorMarkovRandomFields(MRFs),anddirectedGMsarealsoknownasBayesianNetworks(BNs),beliefnetworks,generativemodelsorcausalmodels.1.1DirectedGMs(BayesianNetworks)[4]InBayesianNetworks(B
8、Ns)eachvertexrepresentsarandomvariable,andanarcfromvertexXtovertexY(wealsosaidthatXisoneoftheparentsofY)meansXisoneofthereasonswhyYhappens,i.e.,XcausesY.HenceBNsareacyclic.BNsassumethatavariable