A Scalable Parallel Reinforcement Learning Method Based on Divide-and-Conquer Strategy  

作　　者：YANG Xudong LIU Quan JING Ling LI Jin YANG Kai 

机构地区：[1]Institute of Computer Science and Technology, Soochow University, Suzhou 215006, China [2]Department of Computer Science and Technology, Nanjing University, Nanjing 210093, China

出　　处：《Chinese Journal of Electronics》2013年第2期242-246,共5页电子学报（英文版）

基　　金：This work is supported by the National Natural Science Foundation of China （No.60873116, No.61070223）, Natural Science Foundation of Jiangsu （No.BK2009116）, High School Natural Foundation of Jiangsu （No.09KJA520002）, Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education, Jilin University （No.93K172012K04）.

摘　　要：To conquer the slow convergence and poor scalability problems of reinforcement learning~ a Scalable parallel reinforcement learning method, DCS-SPRL, is proposed on the basis of Divide-and-conquer strategy. In this method, the learning problem with large state space is decomposed into multiple smaller subproblems. Accord- ing to a weighted priority scheduling algorithm, these sub- problems are then dispatched to the learning agents which are able to learn in parallel. Finally, the learning results of each subproblem are merged into a composite solution. The experimental results show that DCS-SPRL has good scalability and needs significantly less computational time.

关 键 词：Parallel reinforcement learning Divide- and-conquer strategy SCALABILITY Large state space. 

分 类 号：TP393[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]