Gaussian Processes and Reinforcement Learning for Identi&ampamp;amp;#64257;cation and.pdf

《Gaussian Processes and Reinforcement Learning for Identi&ampamp;amp;#64257;cation and.pdf》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、GaussianProcessesandReinforcementLearningforIdentificationandControlofanAutonomousBlimpJonathanKo∗DanielJ.Klein†DieterFox∗DirkHaehnel‡∗Dept.ofComputerScience&Engineering,†Dept.ofAeronautics&Astronautics,‡IntelResearchSeattle,UniversityofWashington,UniversityofWashington,Seattle,WASeattle,WASeattle,WA

2、Abstract—Blimpsareapromisingplatformforaerialroboticsandhavebeenstudiedextensivelyforthispurpose.Unlikeotheraerialvehicles,blimpsarerelativelysafeandalsopossesstheabilitytoloiterforlongperiods.Theseadvantages,however,havebeendifficulttoexploitbecauseblimpdynamicsarecomplexandinherentlynon-linear.Thec

3、lassicalapproachtosystemmodelingrepresentsthesystemasanordinarydifferentialequation(ODE)basedonNewtonianprinciples.AmorerecentmodelingapproachisbasedonrepresentingstatetransitionsasaGaussianprocess(GP).Inthispaper,wepresentFig.1.Theleftimageshowstheblimpusedinourtestenvironmentageneraltechniqueforsy

4、stemidentificationthatcombinesequippedwithamotioncapturesystem.Ithasacustomizedgondola(rightthesetwomodelingapproachesintoasingleformulation.Thisisimages)thatincludesanXScalebasedcomputerwithsensors,twoducteddonebytrainingaGaussianprocessontheresidualbetweenthefansthatcanberotatedby360degrees,andaweb

5、cam.non-linearmodelandgroundtruthtrainingdata.TheresultisaGP-enhancedmodelthatprovidesanestimateofuncertaintyinadditiontogivingbetterstatepredictionsthaneitherODEproblemoflearningdynamicmodelsfromtrainingdata[4],orGPalone.WeshowhowtheGP-enhancedmodelcanbe[6].GPshaveseveralkeypropertiesthatmakethemid

6、eallyusedinconjunctionwithreinforcementlearningtogenerateasuitedtoourproblem.Theyarenon-parametric,whichletsblimpcontrollerthatissuperiortothoselearnedwithODEorthemmodelawiderangeofdynamicalsystems.Further-GPmodelsalone.more,theycanautomaticallylearnthesmoothnessandnoiseI.INTRODUCTIONANDMOTIVATIONle

7、velsoftheunderlyingsystem.Finally,theyprovideanotionofuncertaintyaboutthelearnedprocess.ThisuncertaintycanUnmannedaerialvehicles(UAVs)havebecomeahelpfulbeveryvaluablewhenlearningacontroller.componentf