基于半自治agent的profit-sharing增强学习方法研究  

Research of profit-sharing reinforcement learning method based on semi-autonomous agent

作　　者：杨克巍[1] 张少丁[1] 岑凯辉[1] 谭跃进[1] 

机构地区：[1]国防科技大学信息系统与管理学院,长沙410073

出　　处：《计算机工程与应用》2007年第15期72-75,97,共5页Computer Engineering and Applications

基　　金：国家部委"十五"预研项目(the Pre- Research Project of the "Tenth Five- Year- Plan"of China) 。

摘　　要：在基于半自治agent的系统中应用profit-sharing增强学习方法,并与基于动态规划的Q-learning增强学习方法进行比较,在不确定因素较多的动态环境中,当系统状态变化不是一个马尔科夫过程时profit-sharing方法具有很大优势。根据半自治agent中半自治的特性——受制性,提出了一种面向基于半自治agent的增强学习模型,以战场仿真中安全隐蔽的寻找模型为实例对基于半自治agent的profit-sharing增强学习模型进行了试验分析。We exert the profit-sharing reinforcement learning method into the semi-autonomous agent system,and compare it with the other reinforce learning method Q-learning.Profit-sharing method is more robust and fit for the dynamic environment which includes many uncertain factors,especially in the partial MDPs（Markov Decision Processes） environment.Facing the semi - autonomous property of the agent,we propose an improving learning method of profit-sharing in the semi-autonomous agent system and test it in a combat simulation environment that finds the safety hidden space in battlefield.At last we contract and analyze these methods to the others.

关 键 词：增强学习 半自治agent PROFIT-SHARING Q-LEARNING 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]