A CBR-Based Multiagent Reinforcement Learning Approach

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.