摘要
提出了一个新的效用聚类激励学习算法U-Clustering。该算法完全不用像U-Tree算法那样进行边缘节点的生成和测试,它首先根据实例链的观测动作值对实例进行聚类,然后对每个聚类进行特征选择,最后再进行特征压缩,经过压缩后的新特征就成为新的状态空间树节点。通过对NewYorkDriving[2,13]的仿真和算法的实验分析,表明U-Clustering算法对解决大型部分可观测环境问题是比较有效的算法。
That presented in this paper is a new utility clustering based reinforcement learning algorithm called U-Clustering.Unlike the U-Tree,it does not use fringe and related statistical test at all.The U-Clustering algorithm groups the instances that have matching history up to a certain length into a cluster based on the observation-action value of them,and makes the feature selecting and feature compressing for each cluster.The new features become new nodes in the agent＇s internal state space tree.Experimental results in a difficult partially observable driving task called New York Driving show that the U-Clustering algorithm is an effective one for solving the large partially observable problems.
出处
《计算机工程与应用》
CSCD
北大核心
2005年第26期37-42,74,共7页
Computer Engineering and Applications
基金
国家自然科学基金(编号:60075019)资助
关键词
激励学习
效用聚类
部分可观测Markov决策过程
reinforcement learning, utility clustering, partially observable Markov decision processes （POMDPs）
