Abstract: Many multi-Agent Q-learning problems can not be solved because the number of joint actions is exponential in the number of Agents, rendering this approach infeasible for most problems. This paper investigates a regional cooperative of the Q-function by only considering the joint actions in those states in which coordination is actually required. In all other states single-Agent Q-learning is applied. This paper offers a compact state-action value representation, without compromising much in terms of solution quality. It performs experiments in RoboCup-simulation 2D which is the ideal testing platform of multi-agent systems and compared the algorithm to other multi-Agent reinforcement learning algorithms with promising results.

Key words: Markov Decision Processes(MDP), Q-learning, regional cooperative, simulation 2D

摘要： 针对多智能体Q-学习中存在的联合动作指数级增长问题，采用一种局部合作的Q-学习方法，在智能体之间有协作时才考察联合动作，否则只进行简单的个体智能体的Q-学习，从而减少学习时所要考察的状态-动作对值。在机器人足球仿真2D平台上进行的实验表明，该方法比常用多智能体强化学习技术具有更高的效率。

关键词: 马尔可夫决策, Q-学习, 局部合作, 仿真2D

CLC Number: 

• TP311