多agent协同强化学习算法SE-MACOL及其应用

Multi-agent Cooperative Reinforcement Learning Algorithm SE-MACOL and Its Application

针对多agent团队中各成员之间是协作关系且自主决策的学习模型,在此对Q学习算法进行了适当扩充,提出了适合于多agent团队的一种共享经验元组的多agent协同强化学习算法。其中采用新的状态行为的知识表示方法,采用相似性变换和经验元组的共享,能够提高多agent团队协同工作的效率。最后将该算法应用于猎人捕物问题域,实验结果表明该算法能够明显加快多个猎人合作抓捕猎物的进程。

This paper extends Q-learning algorithm properly to multi-agent cooperative team domain, in which members make their decisions independently,and proposes a shared experience tuples multi-agent cooperative reinforcement learning algorithm. A new knowledge representation form composed of sequential pair as （state-value,action-value） is proposed, and experience tuples are shared with other agents in one team by through similarity transformation according to homogeneous subtasks. By importing this learning algorithm,not only the space of state-action is reduced,but also the learning efficiency is improved ,and it shows that cooperation efficiency of the team is improved obviously. In the end ,the algorithm is applied to pursuit game domain,and the result shows the validity of the algorithm that it can speed up the progress of pursuit task.