欢迎来到天天文库
浏览记录
ID:34486912
大小:717.77 KB
页数:6页
时间:2019-03-06
《adaptive robot learning in a non-stationary environmentnew》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、Adaptiverobotlearninginanon-stationaryenvironmentKaryFrämlingHelsinkiUniversityofTechnology,DepartmentofComputerScience,FI-02015HUT,FinlandKary.Framling@hut.fiAbstract.Adaptivecontrolischallenginginreal-worldapplicationssuchasrobotics.Learninghastoberapidenoughtobeperformedinrealtimeandtoavo
2、iddamagetotherobot.Modelsusinglinearfunctionapproximationareinterestinginsuchtasksbecausetheyofferrapidlearningandhavesmallmemoryandprocessingrequirements.Thismakesthemsuitableasadaptivecontrollersinnon-stationaryenvironments,especiallywhenthecontrollerneedstobeanembeddedsystem.Experimentswi
3、thalight-seekingrobotillustratehowtherobotadaptstotheenvironmentbyReinforcementLearningwheretherobotcollectstrainingsamplesbyexploringtheenvironment.1IntroductionTheuseofmachinelearninginreal-worldcontrolapplicationsischallenging.Real-worldtasks,suchasthoseusingrealrobots,involvenoisecomingf
4、romsensors,non-deterministicactionsanduncontrollablechangesintheenvironment.Inrobotics,learningmustberelativelyrapidandpossibletoperformwithoutcausingdamagetotherobot.Onlyinformationthatisavailablefromrobotsensorscanbeusedforlearning.Thismeansthatthelearningmethodshavetobeabletohandlepartial
5、lymissinginformationandsensornoise,whichmaybedifficulttotakeintoaccountinsimulatedenvironments.Artificialneuralnetworks(ANN)areawell-knowntechniqueformachinelearninginnoisyenvironments.Inrealroboticsapplications,however,ANNlearningmaybecometooslowtobepractical.One-layerlinearfunctionapproxim
6、ationANNs(oftencalledAdalines[7])offerfastertrainingthannon-linearANNsandtheirconvergencetoanoptimalsolutioncanusuallybeguaranteed.Thesearepropertiesthatareparticularlyusefulinnon-stationaryenvironmentsthatrequirerapidadaptation,especiallyiftherobothastoexploretheenvironmentandcollecttrainin
7、gsamplesbyitself.Learningbyautonomousexplorationoftheenvironmentisoftenperformedusingreinforcementlearning(RL)methods.Finally,thelimitedmemory-andcomputingpowerneedsofAdalinesmakethemeasytouseinembeddedsystems.Thestructureofthispaperisasfollows.Sec
此文档下载收益归作者所有