计算机工程与应用 ›› 2008, Vol. 44 ›› Issue (24): 156-158.DOI: 10.3778/j.issn.1002-8331.2008.24.047

• 数据库、信号与信息处理 • 上一篇    下一篇

基于先验知识的改进强化学习及其在MAS中应用

毛俊杰,刘国栋   

  1. 江南大学 通信与控制工程学院,江苏 无锡 214122
  • 收稿日期:2007-10-22 修回日期:2008-01-22 出版日期:2008-08-21 发布日期:2008-08-21
  • 通讯作者: 毛俊杰

Modified reinforcement learning based on experience konwledge and its application in MAS

MAO Jun-jie,LIU Guo-dong   

  1. School of Communications and Control Engineering,Jiangnan University,Wuxi,Jiangsu 214122,China
  • Received:2007-10-22 Revised:2008-01-22 Online:2008-08-21 Published:2008-08-21
  • Contact: MAO Jun-jie

摘要: 针对传统的多Agent强化学习算法中,Agent学习效率低的问题,在传统强化学习算法中加入具有经验知识的函数;从心理学角度引入内部激励的概念,并将其作为强化学习的激励信号,与外部激励信号一同作用于强化学习的整个过程。最后将此算法运用到RoboCup仿真中,仿真结果表明该算法的学习效率和收敛速度明显优于传统的强化学习。

关键词: 多智能体系统, 先验知识, 内在激励, 强化学习

Abstract: In order to increase the speed of the agent learning,which is deficient in the triditional reinforcement learing in MAS.The experience konwledge is used in it,and the conception of intrinsic motivation from psychology is introduced.The intrinsic reinforcement,together with extrinsic reinforcement signal act on the whole process of the learning.At last,this algorithm is used for RoboCup simulation,the results of experiment show that the modified algorithm has faster speed to converge and better performance.

Key words: Multi-Agent System(MAS), experience konwledge, intrinsic motivation, reinforcement learning