计算机工程与应用 ›› 2010, Vol. 46 ›› Issue (17): 8-10.DOI: 10.3778/j.issn.1002-8331.2010.17.003

• 博士论坛 • 上一篇    下一篇

集体理性约束的Agent协作强化学习

吴士泓,李德华,潘 莹   

  1. 华中科技大学 图像识别与人工智能研究所,武汉 430074
  • 收稿日期:2010-02-04 修回日期:2010-04-27 出版日期:2010-06-11 发布日期:2010-06-11
  • 通讯作者: 吴士泓

Multi-Agent cooperative reinforcement learning based on collective rationality

WU Shi-hong,LI De-hua,PAN Ying   

  1. Institute for Pattern Recognition and Artificial Intelligence,Huazhong University of Science and Technology,Wuhan 430074,China
  • Received:2010-02-04 Revised:2010-04-27 Online:2010-06-11 Published:2010-06-11
  • Contact: WU Shi-hong

摘要: 将多Agent协作学习过程看作是一个个的阶段博弈,针对博弈中存在多个均衡解的问题,提出一种集体理性约束下的多Agent协作强化学习算法。该算法使得系统中的每个Agent均按照集体利益最大化的集体理性原则进行行为选择,从而解决均衡解一致问题,同时使得集体长期回报值最大化,加快了学习速度。在集体理性的基础上通过评价各Agent对整体任务求解的贡献度,解决信度分配问题。追捕问题的仿真实验结果验证了算法的有效性。

关键词: 多Agent系统, 强化学习, 集体理性

Abstract: The process of multi-Agent cooperative reinforcement learning can be structured as stage games.To solve the problem of equilibrium consistent in the game,an algorithm based on collective rationality is proposed.In the constraint of collective rationality,all the Agents in the system select action in the principle of maximum collective interest,and then acquire the maximum collective interest to improve cooperative learning speed.Moreover,the proposed algorithm measures the contribution of each Agent to the whole task based on the collective rationality to solve the credit assignment problem.The experimental results of the pursuit problem show the efficiency of the proposed algorithm.

Key words: multi-Agent systems, reinforcement learning, collective rationality

中图分类号: