Learned graphical models for probabilistic planning provide a new class of movement primitives
Publikationen: Beitrag in Fachzeitschrift › Artikel › Forschung › (peer-reviewed)
Autoren
Externe Organisationseinheiten
- Technische Universität Graz
- Freie Universität Berlin
Abstract
Biologicalmovementgenerationcombinesthreeinterestingaspects:itsmodularorganizationinmovementprimitives(MPs),itscharacteristicsofstochasticoptimalityunderperturbations,anditsefficiencyintermsoflearning.Acommonapproachtomotorskilllearningistoendowtheprimitiveswithdynamicalsystems.Here,theparametersoftheprimitiveindirectlydefinetheshapeofareferencetrajectory.WeproposeanalternativeMPrepresentationbasedonprobabilisticinferenceinlearnedgraphicalmodelswithnewandinterestingpropertiesthatcomplieswithsalientfeaturesofbiologicalmovementcontrol.Insteadofendowingtheprimitiveswithdynamicalsystems,weproposetoendowMPswithanintrinsicprobabilisticplanningsystem,integratingthepowerofstochasticoptimalcontrol(SOC)methodswithinaMP.Theparameterizationoftheprimitiveisagraphicalmodelthatrepresentsthedynamicsandintrinsiccostfunctionsuchthatinferenceinthisgraphicalmodelyieldsthecontrolpolicy.Weparameterizetheintrinsiccostfunctionusingtask-relevantfeatures,suchastheimportanceofpassingthroughcertainvia-points.Thesystemdynamicsaswellasintrinsiccostfunctionparametersarelearnedinareinforcementlearning(RL)setting.Weevaluateourapproachonacomplex4-linkbalancingtask.Ourexperimentsshowthatourmovementrepresentationfacilitateslearningsignificantlyandleadstobettergeneralizationtonewtasksettingswithoutre-learning.
Details
| Originalsprache | Englisch |
|---|---|
| Aufsatznummer | 97 |
| Seitenumfang | 20 |
| Fachzeitschrift | Frontiers in computational neuroscience |
| Jahrgang | 6.2013 |
| Ausgabenummer | January |
| DOIs | |
| Status | Elektronische Veröffentlichung vor Drucklegung. - 2 Jan. 2013 |
| Extern publiziert | Ja |
