Learned graphical models for probabilistic planning provide a new class of movement primitives
Research output: Contribution to journal › Article › Research › peer-review
Authors
External Organisational units
- 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
| Original language | English |
|---|---|
| Article number | 97 |
| Number of pages | 20 |
| Journal | Frontiers in computational neuroscience |
| Volume | 6.2013 |
| Issue number | January |
| DOIs | |
| Publication status | E-pub ahead of print - 2 Jan 2013 |
| Externally published | Yes |
