基于不平衡学习的分类器博弈模型及其在中国象棋中的应用

Modeling Chess Strategy by Classifier Based on Imbalance Learning and Application in Computer Chinese Chess

计算机博弈是人工智能领域中的热点研究课题.传统计算机博弈模型使用极大极小搜索与评估函数相结合的方式,棋力高低依赖于搜索的深度.在计算性能较低的平台上,搜索深度加深会延长反应时间.因此,提出了一种应用不平衡学习技术使用专家谱训练分类器的机器博弈解决方案,反应时间只相当于一层搜索,且更能体现学习的特性.使用3种经典的不平衡学习方法训练神经网络,并对结果进行了比较.验证了使用分类器模拟中国象棋策略的可能性,以及不平衡学习技术在该策略建模过程中起到的关键作用.

Computer chess game（CCG） is an important topic in the field of artificial intelligence.This technique is widely used in some entertainment PC games and chess games on different platforms.Most CCG systems are developed based on the combination of game tree searching and evaluation functions.When using game tree searching method,the level of the computer player depends on the searching depth.However,deep game tree searching is time-consuming when the games are applied on some mobile platforms such as mobile phone and PDA.In this paper,a novel method is proposed which models Chinese chess strategy by training a classifier.When playing chess games,the trained classifier is used to predict good successor positions for computer player.The training procedure is based on imbalance learning and it uses Chinese chess game records as the training sets.Specifically,the training sets extracted from game records are imbalanced;therefore,imbalance learning methods are employed to modify the original training sets.Compared with the classical CCG system,this new method is as fast as 1-level game tree search when playing games,and it contains an offline learning process.Experimental results demonstrate that the proposed method is able to model Chinese chess strategies and the imbalance learning plays an important role in the modeling process.