一种基于案例推理的多agent强化学习方法研究  

A CBR-Based Multiagent Reinforcement Learning Approach

作　　者：李珺[1] 潘启树[1] 洪炳镕[1] 

机构地区：[1]哈尔滨工业大学计算机科学与技术学院,黑龙江哈尔滨150001

出　　处：《机器人》2009年第4期320-326,共7页Robot

基　　金：国家863计划资助项目(2002AA735041)

摘　　要：提出一种基于案例推理的多agent强化学习方法。构建了系统策略案例库,通过判断agent之间的协作关系选择相应案例库子集。利用模拟退火方法从中寻找最合适的可再用案例策略,agent按照案例指导执行动作选择。在没有可用案例的情况下,agent执行联合行为学习(JAL)。在学习结果的基础上实时更新系统策略案例库。追捕问题的仿真结果表明所提方法明显提高了学习速度与收敛性。A multiagent reinforcement learning approach based on CBR （case-based reasoning） is proposed. The system policy case library is built, and the relevant policy case subset is chosen by judging the cooperation relationship between the agents. Simulated annealing is used to find the fittest and reuseful case policy, and then the agents choose their actions based on the case. And if there is no practicable case in the case library, the agents carry out joint action learning （JAL）. The system policy case library can be updated in real time based on the learning results. The detailed simulation results on pursuit problem are presented to show the superiority of the presented method in learning speed and convergency.

关 键 词：多AGENT强化学习 Q学习 策略再用 基于案例的推理 追捕问题 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]