基于先验知识的强化学习系统  

Prior Knowledge Based Reinforcement Learning System

作　　者：李伟[1] 何雪松[1] 叶庆泰[1] 朱昌明[1] 

机构地区：[1]上海交通大学机械与动力工程学院,上海200030

出　　处：《上海交通大学学报》2004年第8期1362-1365,共4页Journal of Shanghai Jiaotong University

基　　金：国家自然科学基金资助项目(69975013)

摘　　要：针对强化学习算法收敛速度慢的主要原因为强化学习算法所用模型通常都假设系统参数未知、先验知识未知,由此造成该算法从没有任何基础开始搜索最优策略,搜索范围大的问题,提出将强化学习系统建立在先验知识的基础上,既有效利用了前人的工作成果,又加快了算法的收敛速度.通过解决电梯群控问题验证了所提出系统的合理性和有效性.The slow rate of convergence is the main disadvantage of the reinforcement learning (RL) algorithm, which results from that the reinforcement learning model is based on model-free and always assuming any prior knowledge unknown. This makes the algorithm must devise a new policy from scratch. This paper presented a RL system based on prior knowledge relevant to the system, which avoids the duplication of these efforts and accelerates the convergence. An elevator domain knowledge based RL system is developed and it has a good performance.

关 键 词：先验知识 强化学习 电梯 控制系统 

分 类 号：TP182[自动化与计算机技术—控制理论与控制工程]