Abstract: RoboCup（Robot World Cup） is the largest scale robot soccer game, including software simulation and hardware entities from two categories project competition. As an important part of software simulation project, RoboCup simulation 2D has become an outstanding experiment platform in which artificial intelligence and multi-agent cooperation are studied. This paper applies the Q-Learning to RoboCup simulation 2D match local attacking decision, through the introduction of zoning, incentive functions based zoning and decision making for real soccer game action simulation, after training a large number of cycles of learning, making the Agent do the independent action decision, thereby strengthening the multi-agent attacking strength.

Key words: Q-Learning, RoboCup, multi-agent cooperation

摘要： RoboCup是世界上规模最大的机器人足球大赛，包括软件仿真与硬件实体两类项目的比赛。RoboCup仿真2D作为软件仿真项目的重要组成部分，成为研究人工智能和多Agent智能体协作的优秀实验平台。将Q学习应用到RoboCup仿真2D比赛的前场进攻动作决策中，通过引入区域划分，基于区域划分的奖惩函数和对真人足球赛中动作决策的模拟，在经过大量周期的学习训练后，使Agent能够进行自主动作决策，从而加强了多Agent的前场进攻实力。

关键词: Q学习, RoboCup, 多智能体协作