Scaling Techniques for Large Markov Decision Process Planning Problems

  • 格式:pdf
  • 大小:57.13 KB
  • 文档页数:2

ScalingTechniquesforLargeMarkovDecisionProcessPlan-ningProblems

TerranLane&LesliePackKaelblingArtificialIntelligenceLaboratoryMassachusettsInstituteofTechnologyCambridge,Massachusetts02139

http://www.ai.mit.edu@ MIT

PlanninginLargeDomains:TheMarkovdecisionprocess(MDP)formalismhasemergedasapowerfulrepresen-tationforcontrolandplanningdomainsthataresubjecttostochasticeffects.Inparticular,MDPsmodelsituationsinwhichanagentcanexactlyobserveallrelevantaspectsoftheworld’sstatebutinwhichtheeffectsoftheagent’sactionsarenondeterministic.ThoughthetheoryofMDPsiswelldevelopedandexactplanningalgorithmsareknown[3],thesemethodsdonotscaletotheexponentiallylargestatespacesthatarecommonlyofinterestinAIproblems.Inthisproject,weareexaminingapproachestoreducingthecomplexityofMDPplanningtechniquesinsuchlargestatespaceswithanemphasisonclassesofproblemsthatariseinmobileroboticsapplications.

MarkovDecisionProcessesforRoboticPlanning:AMarkovdecisionprocessdescribesasynchronouscontrolprocesswithfourcomponents:astatespacethatspecifiesallpossibleconfigurationsofthesystem(e.g.,thepositionofarobotinamap),anactionspacethatliststheprimitiveactionsavailabletothecontroller(e.g.,arobot’smovementandmanipulationcommands),atransitionfunctionthatspecifiesthe,possiblystochastic,outcomesoftakingeachactioninanystate,andarewardfunctionthatdefinesthegoalsofanagentinthisspace.Inpractice,thestatespaceisoftenexpressedasacross-productofstatevariables(e.g.,xandycoordinatesoralistofwhattheagentisholding),yieldingaspacethatisexponentiallylargeinthenumberofsuchvariables.Itisassumedthatthestatespaceisbothcomplete,inthatallrelevantvariablesareencoded,andfullyobservable,inthattheagentcandirectlysenseallstatevariables.

ThegoalofplanningintheMDPframeworkis,givenacompletedescriptionofanMDP,todevelopanoptimalpolicythatspecifieswhichactiontotakeinanystate.Anoptimalpolicyisonethatmaximizesexpectedaccumulatedrewardoverthelifetimeoftheagent.WhileitisoftenpossibletocompactlydescribeevenlargeMDPswithdynamicBayesnetworkrepresentationsofthetransitionfunction[1],afulldescriptionofthepolicymaystillrequireacompleteenumerationofthestatespace.

Inmobilerobotapplications,wecanformulateMDPsthatrepresentthedynamicsoftherobot’senvironmentaswellasthetasksthattherobotistoaccomplish.PartofanexampletargetproblemisdisplayedinFigure1,whichshowsalow-resolutiondiscretemapofoneofthetenfloorsofthebuildinghousingtheMITAILab.Inthistask,therobotisrequiredtodeliverpackagestovariousofficesinthebuildingwhilemaintainingitsbatterylevelandperiodicallycheckingfornewlyarrivedmailinthedropneartheelevators.Thetaskisfurthercomplicatedbydoorsthatmaybelockedatdifferenttimesofthedayandofficeresidentswhomaybeavailableonlysporadically.Evenaconservativerepresentationofthisproblemrequireswellinexcessof2500statestofullyrepresent.

•10floors•>1800locationsperfloor•∼400deliverylocations•25batterylevels•∼450doors(somelockable)

Figure1:TargetDomain55ApproachestoScalablePlanning:WhileanumberofapproachestoplanninginexponentiallylargeMDPshavebeenproposed,nosingletechniquehasprovenpowerfulenoughtocompletelyaddressproblemsofthescalede-scribedabove.Instead,weseektoattacklargestatespacesbyintegratinganumberofexistingtechniqueswithourownnovelmethods.Currently,wearedevelopingplanningsystemsbasedonthefollowingmethods:

HierarchicalStateDecompositionStatespacescanbedecomposedintogeographicregions,suchasfloors,hall-ways,orrooms,andplanningcarriedoutlocallyratherthanglobally.Thelocalplans,ormacros(representedinPrecup’s“options”framework[2]),canthenbeusedas“meta-actions”inahigher-levelplanningoperationtoproduceaglobalplan.

RewardFunctionDecompositionInmanyMDPproblemdomains,thedynamicsoftheenvironment(expressedviathestatespaceandtransitionfunction)arefixedacrossproblem-solvingepisodes,butthetasksorgoals(expressedbytherewardfunction)aredynamic.Wewouldthusliketoamortizeplanningeffortacrossmul-tipleepisodesbydevelopingsub-plansthataccountforenvironmentalstructurebutwhichcanbereusedinresponsetocurrentgoalvalues.Thisleadstoanalternatetypeofdecompositionthatgeneratesmacrosforachievingsub-goalsoftheglobalrewardfunction.Suchmacrosmayalsoyieldasubstantialimprovementinone-shotplanningeffort:sub-goalsareoftenexplicitlyassociatedwithparticularvariablesofthestatespace(e.g.,abitmaybeusedtoindicatethataparticularpackagehasbeendelivered)anddiscardingvariablesirrelevanttothesub-goalcanproduceanexponentialimprovementinplanningeffortforthattask.Byapply-ingthisprocesstoeachsub-goal,weproduceasetofmacrosforhandlingpartsoftheglobaltask.Again,ahigher-levelplanningprocessisnecessarytointegratethemacrosintoaglobalplan.