基于强化学习的多智能体动态协作规则提取  

EXTRACTING DYNAMIC COOPERATIVE RULES BETWEEN MULTIAGENT USING REINFORCEMENT LEARNING

作　　者：李晓萌[1] 杨煜普[1] 许晓鸣[1] 

机构地区：[1]上海交通大学自动化研究所,上海200030

出　　处：《模式识别与人工智能》2002年第4期463-467,共5页Pattern Recognition and Artificial Intelligence

摘　　要：在分布式的动态环境下,多智能体系统的协作是建立在规则集合上的动态过程,因此需要建立动态的协作规则.多智能体强化学习的平稳状态本质上即是智能体之间的协作规则,据此提出一种基于强化学习的协作规则提取的方法,并由此构成智能体决策的新结构,最后用实例进行分析和证明了所提出的方法与单纯的强化学习方法相比较,具有如下优点:1)提取的规则可以加快多智能体的协作决策过程;2)规则的动态变化可以适应环境的动态变化;3)规则可以避免多次重复的学习过程.Cooperative process in a multiagent system should be real-time in the distributed dynamic environment. Usually, cooperative process in a multiagent system is a dynamic process built on the rule set, in which rules also change dynamically. Equilibriums of multiple agents' interactive and reinforcement learning process are cooperative rules essentially. Recognized on this, a method of extracting dynamic cooperative rules is proposed. The method is based on single agent' s reinforcement learning and two-layer rule extraction model, and a new structure of agents' interactive decision based on belief function is built. The case study proofs that the method has three advantages comparing with the pure reinforcement learning: 1)The extracted rules can speed up the process of multi-agent cooperative decision; 2)The rules can change dynamically for adapting to the environment change; 3)The rules could avoid many repetitive learning processes.

关 键 词：强化学习 多智能体 动态协作 规则提取 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]