【摘要】 针对多agent系统强化学习中,状态空间和动作空间随着agent个数的增加成指数倍增长,进而导致维数灾难、学习速度慢和收敛性差的问题,提出了一种新型的混合强化学习方法,用于改进传统的多agent协作强化学习;该算法基于Friend-or-Foe Q-学习,事先采用聚类分析法对状态空间和动作空间进行预处理,降低空间维数后再进行强化学习,这就避免了同等状态环境下的重复劳动和对动作集的盲目搜索,理论上大大提高了agent的学习速度和算法的收敛性;文章首先进行改进算法的思想概述,然后给出了改进算法的学习框架和算法的一般描述。更多还原

【Abstract】 As State space and Action space grow exponentially in Multi-agent Cooperative Reinforcement Learning method,which leads to dimension disaster and the slow learning speed and poor convergence,a new hybrid Reinforcement Learning approach is proposed in this paper to solve the problem of above.This algorithm is based on Friend-or-Foe Q-Learning.It doesn’t do reinforcement learning until the clustering analysis is used to pre-treat both State space and Action space,to reduce the space dimension.This advanced approach not only avoids the duplication of work and blind search of the action set but also enhances the learning speed of the agents and the convergence of the algorithm greatly.This paper first outlines the idea of this advanced algorithm,and then gives the framework and the general description.更多还原