基于云端可视化交互的强化学习平台  

Reinforcement Learning Platform Based on Cloud Visual Interaction

作　　者：姚铁锤 王珏[1,2] 王彦棡 迟学斌[1,2] 王晓光 YAO Tiechui;WANG Jue;WANG Yangang;CHI Xuebin;WANG Xiaoguang(Computer Network Information Center,Chinese Academy of Sciences,Beijing 100190,China;School of Computer Science and Technology,University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院计算机网络信息中心,北京100190 [2]中国科学院大学计算机科学与技术学院,北京100049

出　　处：《计算机工程》2021年第5期316-320,共5页Computer Engineering

基　　金：国家重点研发计划“大规模并行计算的工具库和领域相关基础软件包”(2017YFB0202202);“中国科技云”建设工程(二期)项目“超算资源池建设”(XXH13503);国家电网有限公司总部科技项目“电力人工智能实验及公共服务平台技术”(SGGR0000JSJS1800569)。

摘　　要：强化学习是一个与环境交互的学习过程,在实验场景中,训练环境部署的可扩展性和算法验证的便捷性常受限于物理引擎和渲染模块的高耦合性。为对物理引擎和渲染模块进行解耦,构建一种面向物理引擎和渲染模块的云端交互式模型,其中包括操作字典、元素字典和对应的算法接口,并基于该模型实现模拟器。通过集成模拟器、可视化工具和知识管理等组件,搭建支持云端可视化交互的强化学习平台,并以MuJoCo物理引擎为例,验证Web模拟器接入自定义物理引擎的便捷性。实验和分析结果验证了该模型的有效性,其可方便接入平台,实现云端渲染并提高所属集群的利用率。Reinforcement learning is a learning process that interacts with the environment.In the experiment environment,the scalability of the training environment deployment and the convenience of algorithm verification are often limited by the high coupling between the physics engine and the rendering module.To solve the problem,this paper proposes a Cloud Interactive Model(CIM)for physics engine and rendering module,which consists of an operation dictionary,element dictionary and relevant algorithm interfaces,and on this basis implements a simulator.Furthermore,this paper integrates the simulator,visualization tools,knowledge management and other components to build a Reinforcement Learning Platform(RLP)supporting cloud visual interaction.By taking the MuJoCo physics engine as an example,the Web simulator is verified for its convenience of access to a custom physics engine.Experimental and analytical results show that this model can be conveniently connected to the platform to realize cloud rendering and improve the utilization rate of its cluster.

关 键 词：强化学习平台 物理引擎 渲染模块 云端可视化交互 接口标准 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]