资源描述:
《Abstraction of State-Action Space Utilizing.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、AbstractionofState-ActionSpaceUtilizingPropertiesoftheBodyandtheEnvironment-Applicationtoa3-dimensionalsnake-likerobotthatoperatesonrubble-KazuyukiITOSoKUROEToshiharuKOBAYASHIHOSEIUniv.HOSEIUniv.MitsubishiChemicalCorp.Tokyo,JapanTokyo,JapanTokyo,Japanito@hosei.ac.jpso.kuroe.4
2、b@stu.hosei.ac.jp3619187@cc.m-kagaku.co.jpAbstract—Inthispaper,weaddresstheautonomouscontrolofaIntheapproachesthatbelongtothelattercategory,several3-dimensionalsnake-likerobotbyusingreinforcementlearning,techniquesareconsidered,e.g.,neuralnetworks[2–3],fuzzyandweapplyitinthec
3、aseofrubble.Ingeneral,snake-likelogic[4–6],orastochasticapproach[7–8]isemployedforrobotshavehighmobilitythatisrealizedbymanydegreesofgeneralizingthestate-actionspace.Acombinationofthesefreedom,andtheycanmoveonrubble.However,themanytechniquesmayresultindecreasedsizeofthestate-
4、actiondegreesoffreedomcausethestateexplosionproblem,andthespacesandreducedreinforcementlearningload;however,complexityoftherubbleresultsinincompletelearning.Therefore,itisimpossibletoapplyreinforcementlearningtotheseapproachesstillhaveproblems.Ingeneral,moretimeisconventional
5、snake-likerobotsthatmoveonrubble.Inthispaper,requiredforgeneralization.Asaresult,thetotallearningtime,tosolvetheseproblems,wefocusonpropertiesoftherealincludingthetimetakenforreinforcementlearningandenvironmentandthedynamicsofamechanicalbody.Wedesigngeneralization,isstilltool
6、ongtobeofpracticaluse.thebodyoftherobotforabstractingthenecessarysmallstate-Incontrast,intherealworld,higherorganismslearninactionspacebyconsideringreal-worldproperties,andwemakerealtimebytrialanderror,andtheacquiredlearninghasitpossibletoapplyreinforcementlearning.Todemonstr
7、atethegeneralapplication.Thereasonbehindsuchorganisms’effectivenessoftheproposedsnake-likerobot,weconductedlearningabilityisstilldebated.Inembodiedcognitivescienceexperimentswherelearningwascompletedwithinreasonable[9],orecologicalpsychology[10],thebodyisconsideredtotimeandth
8、eroboteffectivelyadapteditselftoanunknown3-dimensionalenvironment.pl