資源描述：
《On-line Optimization of Sequential Monte Carlo Methods 》由會員上傳分享，免費在線閱讀，更多相關(guān)內(nèi)容在學(xué)術(shù)論文-天天文庫。
1、ProceedingsoftheAmericanControlConferenceAnchorage,AKMay8-10.2002On-lineOptimizationofSequentialMonteCarloMethodsusingStochasticApproximationArnaudDoucet’,VladislavB.TadiCDepartmentofElectricalandElectronicEngineering,TheUniversityofMelbourne,Parkville,Victoria3052,Australia.Email:{a.doucet,
2、v.tadic}Qee.mu.oz.auAbstractderweakassumptions,itcanbetypicallyshownthatthesealgorithmsconvergeinacertainsensetowardstheSequentialMonteCarlo(SMC)methodsakaParticleposteriorprobabilitydistributionsofinterestasymp-ateringtechniquesareasetofpowerfulandversatiletoticallyinthenumberofparticles[5]
3、,[SI.However,simulation-basedmethodstoperformoptimalstatees-theperformanceofSMCalgorithmsdependsheavilyontimationinnon-linearnon-Gaussianstatespacemodelsthevariousparametersofthealgorithms.Considerfor[SI.Inthisapproach,theposteriorprobabilitydistri-exampletheclassofSequentialImportanceSampli
4、ngbutionsofinterestareestimatedusingacloudofran-Resampling(SISR)algorithms[7].Currentalgorithmsdomsampleswhicharecarriedovertimeusingimpor-aretypicallydesignedsoastooptimizesome“l(fā)ocal”tancesamplingandresamplingtechniques.Currental-criteriasuchastheconditionalvarianceoftheimpor-gorithmsaretyp
5、icallydesignedsoastooptimizesometanceweightsintheimportancesamplingsteporthe“l(fā)ocal1’criteriasuchastheconditionalvarianceoftheconditionalvarianceofthenumberofoffspringintheimportanceweightsintheimportancesamplingstep.resamplingstep.However,theeffectoftheselocalopti-However,theeffectoftheseloc
6、aloptimizationsisnotmizationsisnotclearontheglobalperformanceoftheclearontheglobalperformanceofthealgorithm;e.g.algorithm.Forexample,samplingwithanon-locally.samplingwithanon-locallyoptimalimportancedistri-optimalimportancedistributionatagiventimecouldbutionmightbebeneficialatfurthertimestep
7、s.Webebeneficialatfurthertimesteps.Soevenifoptimiz-presenthereanaltemativeprincipledapproachwhereing“l(fā)ocal”criteriaissensible,onewouldpreferinap-theSMCisparametrizedanditsparametersoptimizedplicationstodesignanalgorithmoptimizinga“global”withrespec
