Robocup半场防守中的一种强化学习算法

A Reinforcement Learning Method for Robocup Soccer Half Field Defense

Robocup仿真比赛是研究多Agent之间协作和对抗理论的优秀平台,提高Agent的防守能力是一个具有挑战性的问题。为制定合理的防守策略,将Robocup比赛中的一个子任务——半场防守任务分解为多个一对一防守任务,采用了基于Markov对策的强化学习方法解决这种零和交互问题,给出了具体的学习算法。将该算法应用到3D仿真球队——大连理工大学梦之翼(Fantasia)球队,在实际比赛过程中取得了良好效果。验证了采用Markov零和对策的强化学习算法在一对一防守中优于手工代码的结论。

Robocup soccer simulation is an excellent platform in which colhboration and counterwork among multi - agent are studied. It is a challenging problem to improve agent＇s defense ability. In order to design reasonable defending policy, decompose a subtask, half field defense, into some one- vs-one defense subtask and pose it as a problem of zero-sum Markov games. In this paper, a reinforcement learning method based on Markov game is developed and implemented in 3D simulation soccer team——DUT Fantasia. In real matches, this arithmetic is approved to be efficient and better than manual - coding in one- vs- one defense subtask.