欢迎来到天天文库
浏览记录
ID:40718303
大小:197.80 KB
页数:13页
时间:2019-08-06
《Hidden Markov Models Fundamentals》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、HiddenMarkovModelsFundamentalsDanielRamageCS229SectionNotesDecember1,2007AbstractHowcanweapplymachinelearningtodatathatisrepresentedasasequenceofobservationsovertime?Forinstance,wemightbeinterestedindiscoveringthesequenceofwordsthatsomeonespokebasedonanaudiorecordingoftheirspeech.Or
2、wemightbeinterestedinannotatingasequenceofwordswiththeirpart-of-speechtags.ThesenotesprovidesathoroughmathematicalintroductiontotheconceptofMarkovModelsaformalismforreasoningaboutstatesovertimeandHiddenMarkovModelswherewewishtorecoveraseriesofstatesfromaseriesofobservations.Then
3、alsectionincludessomepointerstoresourcesthatpresentthismaterialfromotherperspectives.1MarkovModelsGivenasetofstatesS=fs1;s2;:::sjSjgwecanobserveaseriesovertime~z2ST.Forexample,wemighthavethestatesfromaweathersystemS=fsun;cloud;raingwithjSj=3andobservetheweatheroverafewdaysfz1=ssun;z
4、2=scloud;z3=scloud;z4=srain;z5=scloudgwithT=5.Theobservedstatesofourweatherexamplerepresenttheoutputofarandomprocessovertime.Withoutsomefurtherassumptions,statesjattimetcouldbeafunctionofanynumberofvariables,includingallthestatesfromtimes1tot 1andpossiblymanyothersthatwedon'tevenmod
5、el.However,wewillmaketwoMarkovassumptionsthatwillallowustotractablyreasonabouttimeseries.Thelimitedhorizonassumptionisthattheprobabilityofbeinginastateattimetdependsonlyonthestateattimet 1.Theintuitionunderlyingthisassumptionisthatthestateattimetrepresentsenoughsummaryofthepasttor
6、easonablypredictthefuture.Formally:P(ztjzt 1;zt 2;:::;z1)=P(ztjzt 1)Thestationaryprocessassumptionisthattheconditionaldistributionovernextstategivencurrentstatedoesnotchangeovertime.Formally:1P(ztjzt 1)=P(z2jz1);t22:::TAsaconvention,wewillalsoassumethatthereisaninitialstateandinitia
7、lobservationz0s0,wheres0representstheinitialprobabilitydistributionoverstatesattime0.Thisnotationalconvenienceallowsustoencodeourbeliefaboutthepriorprobabilityofseeingtherstrealstatez1asP(z1jz0).NotethatP(ztjzt 1;:::;z1)=P(ztjzt 1;:::;z1;z0)becausewe'vedenedz0=s0foranystatesequen
8、ce.(Otherpresentati
此文档下载收益归作者所有