返回
31-Learning Complex Neural Network Policies with Trajectory Optimization(icml2014)

31-Learning Complex Neural Network Policies with Trajectory Optimization(icml2014)

ID：40704475

大小：1.39 MB

页数：10页

时间：2019-08-06

31-Learning Complex Neural Network Policies with Trajectory Optimization(icml2014)_第1页
31-Learning Complex Neural Network Policies with Trajectory Optimization(icml2014)_第2页
31-Learning Complex Neural Network Policies with Trajectory Optimization(icml2014)_第3页
31-Learning Complex Neural Network Policies with Trajectory Optimization(icml2014)_第4页
31-Learning Complex Neural Network Policies with Trajectory Optimization(icml2014)_第5页
资源描述：

《31-Learning Complex Neural Network Policies with Trajectory Optimization(icml2014)》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、LearningComplexNeuralNetworkPolicieswithTrajectoryOptimizationSergeyLevineSVLEVINE@CS.STANFORD.EDUComputerScienceDepartment,StanfordUniversity,Stanford,CA94305USAVladlenKoltunVLADLEN@ADOBE.COMAdobeResearch,SanFrancisco,CA94103USAAbstract2013).Suchspecializedpolicyclassesarelimitedinthetypesofbe

2、haviorstheycanrepresent,andengineeringnewDirectpolicysearchmethodsofferthepromisepolicyclassesrequiresconsiderableeffort.ofautomaticallylearningcontrollersforcom-plex,high-dimensionaltasks.However,priorap-Inrecentwork,weintroducedanewclassofpolicysearchplicationsofpolicysearchoftenrequiredspe-a

3、lgorithmsthatcanlearnmuchmorecomplexpoliciesbycialized,low-dimensionalpolicyclasses,limit-usingmodel-basedtrajectoryoptimizationtoguidethepol-ingtheirgenerality.Inthiswork,weintroduceicysearch(Levine&Koltun,2013a;b).Byoptimizingtra-apolicysearchalgorithmthatcandirectlylearnjectoriesintandemwith

4、thepolicy,guidedpolicysearchhigh-dimensional,general-purposepolicies,rep-methodscombinetheflexibilityoftrajectoryoptimizationresentedbyneuralnetworks.Weformulatethewiththegeneralityofpolicysearch.Thesemethodscanpolicysearchproblemasanoptimizationoverscaletohighlycomplexpolicyclassesandcanbeusedt

5、otrajectorydistributions,alternatingbetweenopti-traingeneral-purposeneuralnetworkcontrollersthatdonotmizingthepolicytomatchthetrajectories,andrequiretask-specificengineering.Furthermore,thetrainingoptimizingthetrajectoriestomatchthepolicytrajectoriescanbeinitializedwithexamplesforlearningandmini

6、mizeexpectedcost.Ourmethodcanfromdemonstration.learnpoliciesforcomplextaskssuchasbipedalAkeychallengeinguidedpolicysearchisensuringthatpushrecoveryandwalkingonuneventerrain,thetrajectoriesareusefulforlearningthepolicy,sincenotwhileoutperformingpriormethods.alltrajectoriescanberealizedbypolicies

7、fromaparticularpolicyclass.Forexample,apolicyprovidedwithpartialobservationscannotmakedecisionsbasedonunobserved1.Introductionstatevariables.Inthispaper,wepresentaconstrainedDirectpolicysearchoffersthepromiseofautomaticallyguidedp

当前文档最多预览五页，下载文档查看全文

此文档下载收益归作者所有

当前文档最多预览五页，下载文档查看全文
温馨提示：
1. 部分包含数学公式或PPT动画的文件，查看预览时可能会显示错乱或异常，文件下载后无此问题，请放心下载。
2. 本文档由用户上传，版权归属用户，天天文库负责整理代发布。如果您对本文档版权有争议请及时联系客服。
3. 下载前请仔细阅读文档内容，确认文档内容符合您的需求后进行下载，若出现内容与标题不符可向本站投诉处理。
4. 下载文档时可能由于网络波动等原因无法下载或下载错误，付费完成后未能成功下载的用户请联系客服处理。