基于最小代价的多分类器动态集成

本文提出一种基于最小代价准则的分类器动态集成方法．与一般方法不同，动态集成是根据“性能预测特征”，动态地为每一样本选择最适合的一组分类器进行集成．该选择基于使误识代价与时间代价最小化的准则，改变代价函数的定义可以方便地达到识别率与识别速度之间的不同折衷．本文中提出了两种分类器动态集成的方法，并介绍了在联机手写汉字识别中的具体应用．在实验中使了３个分类器进行动态集成，因此，得到７种分类组合．在预先定义的代价意义下，我们比较了动态集成方法和其它７种固定方法的性能．实验结果证明了动态集成方法的高灵活性、实用性和提高系统综合性能的能力．

DYNAMIC COMBINATION OF MULTI CLASSIFIERS BASED ON MINIMUM COST CRITERION

This paper presents a dynamic multi classifier combination method based on minimum cost criterion. Different from common methods, dynamic combination scheme selects the most suitable group of classifiers for each input sample according to its Performance Predication Features (PPFs), that is, for different input, use different group of classifiers. The PPFs are those numerical features which have great influence on the performance of classifiers. Because every classifier behaves differently in various area of PPF's space, in each certain area suitable group of classifiers should be uniquely selected. The selection is based on the criterion that the cost caused by recognition error and recognition time should be minimized. Evidently adjustment of cost function will lead to various trade offs between recognition rate and recognition speed. In the paper, two implementation approaches for dynamic classifier combination are presented, and the concrete application in on line handwritten Chinese character recognition is introduced. In the experiment three classifiers are chosen to perform dynamic combination, and thus seven kind of classifier groups can be formed. Under meanings of the pre defined cost, this paper compares the performance of the dynamic combination scheme and other seven fixed schemes. The experimental results demonstrate its merits of high flexibility, practicality, the capability to improve the system's comprehensive performance.