资源描述：
《a semi-supervised method for learning the structure of robot environment interactionsnew》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。
1、ASemi-SupervisedMethodforLearningtheStructureofRobotEnvironmentInteractionsAxelGroßmann1,MatthiasWendt1andJeremyWyatt21DepartmentofComputerScienceTechnischeUniversit¨atDresdenDresden,Germanyfaxg,mw177754g@inf.tu-dresden.de2SchoolofComputerScienceTheUniversityofBirm
2、inghamBirmingham,UK,B152TTjlw@cs.bham.ac.ukAbstract.Foramobilerobottoactautonomously,itmustbeabletoconstructamodelofitsinteractionwiththeenvironment.Oatesetal.developedanunsu-pervisedlearningmethodthatproducesclustersofrobotexperiencesbasedonthedynamicsoftheinterac
3、tion,ratherthanonstaticfeatures.Wepresentasemi-supervisedextensionoftheirtechniquethatusesinformationaboutthecontrollerandthetaskoftherobotto(i)segmentthestreamofexperiences,(ii)optimisethefinalnumberofclustersand(iii)automaticallyselecttheindividualsensorstofeedtot
4、heclusteringprocess.ThetechniqueisevaluatedonaPioneer2robotnavigatingobstaclesandpassingthroughdoorsinanofficeenvironment.Weshowthatthetechniqueisabletoclassifyhighdimensionalrobottimeseriesseveraltimesthelengthpreviouslyhandledwithanaccuracyof91%.1IntroductionWewou
5、ldlikeourmobilerobotstooperatesuccessfullyintherealworld.Indepen-dentlyofwhetherouraimistrulyautonomousbehaviourorjustreliableandrobustoperation,thisrequirestherobotstocollectinformationabouttheinteractionwiththephysicalenvironment.Inparticular,wewantanautomatictec
6、hniqueforconstructingamodeloftheworlddynamics.Sinceourparticulargoalistousesuchamodelforexecutionmonitoring[4,5]atthelevelofreactivecontrol,itshouldsupportpredictionsaboutthequalitativeoutcomeofactionsaswellashelpinexplainingsituationsinwhichtheactionshadunintended
7、effects.Astheinteractionofarobotwiththeenvironmentiscomplex,adescriptionofitwillbedifficulttoobtain.Ontheonehand,wewouldfavouranunsupervisedlearningtechnique,e.g.,theworkbyOatesetal.[14]onclusteringrobot-sensordatausingdynamictimewarpingassimilaritymeasure.Ontheothe
8、rhand,learningaworldmodelisfundamentallyasupervisedlearningproblem.Astheentireworlddynamicswillbehugeinanyrealisticapplication,itwillbeimportantt
