摘要
经典连续U-树算法使用分布检验来确定抽象状态的最佳分裂点,但选取合适的置信阈值非常困难.提出一种基于最优的最佳分裂点选取方法,该方法将抽象状态的最佳分裂点选取问题转化为一个最优问题,从而规避了置信阈值大小难以确定的问题,并从理论上减少了连续U-树算法的时间复杂度.通过消解协商僵局的学习任务实验验证了它的有效性,表明了算法的性能得到增强.
The classical continuous U-tree algorithm employs distribution tests （e. g. Kolmogorov-Smirnov test and information gain ratio test） to determine the best splitting points of abstract states, but it is very difficult to set a confidence threshold properly. A method for selecting the best splitting points of abstract states in continuous U-tree based on optimization is put forward. This method turns the task of selecting the best splitting points into and optimization one. As a result, it avoids the difficulty of setting the appropriate confidence threshold in the classical algorithm and reduces the time complexity of the algorithm in theory. As is shown by the results of experiments upon the complex learning task getting rid of negotiation deadlocks, the method is valid and the performance of the continuous U-tree algorithm utilizing the method is enhanced.
出处
《上海交通大学学报》
EI
CAS
CSCD
北大核心
2008年第2期279-284,共6页
Journal of Shanghai Jiaotong University
基金
广东省自然科学基金(06029281)资助项目
关键词
连续U-树
状态抽象
最佳分裂点
协商僵局
continuous U-tree
state abstraction
best splitting point
negotiation deadlock
