摘要
介绍了一种基于先验知识的强化学习方法,它将传统的规则控制方法和强化学习方法相结合,在保留了已知的部分规则的情况下,利用强化学习方法对基本行为的融合机制进行了完善;同时,利用已知的规则知识对学习器进行指导,保证了学习向正确方向进行,有利于学习收敛速度的提高.文章给出了2种实现方法的结合方式,并给出了学习器的结构及参数和函数设定.最后以机器人围捕为研究背景,实现了移动机器人的自适应式行为融合,并利用仿真实验对其有效性进行验证.结果表明该方法具有收敛快、学习效果好的特点.
A method to reinforce learning based on prior knowledge was proposed, combining the traditional rule control method with the reinforcement learning method. The action fusion mechanism preserves the partially known rules and utilizes the reinforcement learning to accomplish modification of rules. At the same time, the partially known rules give guidance to the learner, which may guarantee the correct learning direction and speed up the convergence. The combination pattern of the two methods is presented, with the architecture and the parameter setting of the learner. The method was used for adaptive action fusion of a mobile robot in a "pursuitevasion" game, and its efficiency was shown by simulation results. The results prove that this method converges in less time and has a good learning result.
出处
《哈尔滨工程大学学报》
EI
CAS
CSCD
北大核心
2005年第5期586-590,613,共6页
Journal of Harbin Engineering University
关键词
强化学习
多机器人
行为融合
reinforcement learning
multi-agent
action fusion