多智能体强化学习及其在足球机器人角色分配中的应用-控制理论与应用.PDFVIP

26 4 Vol. 26 No. 4 2009 4 Control Theory Applications Apr. 2009 2009 , , (1. , 110178; 2. , 110023) : , , , . , . , . , , . : ; ; ; ; : TP242 : A Multi-agent reinforcement learning and its application to role assignment of robot soccer DUAN Yong , CUI Bao-xia , XU Xin-he (1. College of Information Science and Engineering, Shenyang University of Technology, Shenyang Liaoning 110178, China; 2. Institute of AI and Robotics, Northeastern University, Shenyang Liaoning 110023, China) Abstract: Robot soccer is a typical multi-agent system. The action selected by each robot player not only depends on the current field state, but is also impacted by other players. Hence, the decision-making strategy of robot soccer obtained by reinforcement learning needs the information of the joint-state and the joint-action of multiple agents. A multi-agent reinforcement learning method based on the action prediction of agents is proposed. The Naive Bayes classifier is applied to predict the actions of other agents. Moreover, the sharing-policy mechanism is introduced into multi-agent reinforcement learning system for exchanging the learning policies among agents. It can increase the learning speed effectively. Finally, the proposed approach is applied to learn the role assignment strategy in realizing the cooperation and coordination between robots. Key words: multi-agent system; reinforcement learning; naive bayes classifier; robot soccer; role assignment 1 (Introduction)