資源描述：
《Markov chain Monte Carlo in action- a tutorial》由會員上傳分享，免費在線閱讀，更多相關內容在學術論文-天天文庫。
1、MarkovchainMonteCarloinaction:atutorialPeterJ.GreenUniversityofBristol,DepartmentofMathematics,Bristol,BS81TW,UK.P.J.Green@bristol.ac.uk1.IntroductionMarkovchainMonteCarloisprobablyabout50yearsold,andhasbeenbothdevelopedandextensivelyusedinphysicsforthelastfourdecades.However,themostspectacularinc
2、reaseinitsimpactandin
uenceinstatisticsandprobabilityhascomesincethelate'80's.Ithasnowcometobeanall-pervadingtechniqueinstatisticalcomputation,inparticularforBayesianinference,andespeciallyincomplexstochasticsystems.2.Cyclonesexample:pointprocessesandchangepointsWewillillustratetheideasofMCMCwitha
3、runningexample:theobservationsareapointprocessofeventsattimesy;y;:::;yinaninterval[0;L).Wesupposetheeventsoccurasa12NPoissonprocess
4、butatapossiblynon-uniformrate:sayx(t)perunittime,attimet;wewishtomakeinferenceaboutx(t).Weconsideraseriesofmodels,ultimatelyallowinganunknownnumberofchangepoints,unkn
5、ownhyperparameters,andaparametricperiodiccomponent.ThemodelsandtherespectivealgorithmsandinferenceswillbeillustratedbyananalysisofadatasetofthetimesofcycloneshittingtheBayofBengal;therewere141cyclonesoveraperiodof100years.Model1:constantrate.Firstsupposethatx(t)xforallt.Thenthetimesoftheeventsare
6、immaterial:weobserveNeventsinatimeintervaloflengthL;theobviousestimateofxNbisx=,themaximumlikelihoodestimatorofxundertheassumptionthatNhasaPoissonLdistribution,withmeanxL.Model2:constantrate,theBayesianway.ForaBayesianapproachtothisexample,supposethatwehavepriorinformationaboutx(frompreviousstudie
7、s,forexample).Supposewecanmodelthisbyx,( ;):Thenwe ndthataposteriorixhasaGammadistributionwithmean(+N)=(+L),orapproximatelyN=LifNandLarelargecomparedwithand.Thuswithalotofdata,theBayesianposteriormeanisclosetothemaximumlikelihoodestimator.ThereisnoneedforMCMCinthismodel:youcancalculatetheposterio
8、rexactly,andrecogniseitasastandarddistribution;itonlyworkedlikethisbecauseweusedaconjugateprior.Model3:constantrate,withhyperparameter.Supposeyouarereluctanttospecifyyourpriorfully:youarehappytosayx,( ;)andcansp
