学习相关源证据

将两个相关证据看成是一个相关源证据与两个独立源证据正交合成的结果，这两个相关证据的合成就归结为这3个相互独立的源证据的正交合成。利用证据理论组合相关多分类器，每个分类器提供的证据是相关的，利用遗算法从样本中学习最优的相关源证据，基于合成证据确定最终的组合分类结果。字符识别实验结果表明，这种相关证据模型能有效利用分类器之间的相关性，从而能提高多分类器的组合识别性能。

Learning Dependent Original Evidences

Dempster Shafer evidence theory is an important tool for handling uncertainties in artificial intelligence and has many successful applications. Its deficiency is inability to process dependent evidences. The authors proposed an improved model, in which two dependent evidences are thought of as resulting from orthogonal sum of dependent original evidence and two independent lines of original evidence, respectively. The dependent original evidence and independent original evidence represent dependent and independent parts of evidence, respectively. The dependent evidence model can be easily extended to one for N lines of dependent evidence, in which any dependent evidence is thought of as resulted from orthogonal sum of dependent original evidence and corresponding independent original evidence. The problem of combining dependent evidence is reduced to the one of combining dependent original evidence and independent original evidence using Dempster combining formula. For this purpose, the independent original evidence must be identified among dependent evidence and dependent original evidence. This is an inverse problem in evidence theory. Its solution is given. But determining dependent original evidence is a difficult and unsolved problem. In this work, it is optimally determined by Goldberg's simple genetic algorithm. In multiple classifier combination, the recognition rates are taken as the fitness functions. Evidence supplied by component classifier is determined by its recognition rate, error rate, and rejection rate. The multiple classifier combination method is used in handwriting recognition. Three classifiers are all nearest neighbor classifiers. Their features are different. They are Legendre moment, Pseudo Zernike moment, and Zernike moment, respectively. Training set includes 1?000 samples and is used to determine the recognition rate, error rate, and rejection rate of the three component classifiers. Test set also include 1?000 samples and is used to judge the performance of combined multiple classifiers. Recognition results of combined classifiers indicate that combining multiple classifiers using the dependent evidence model give a higher recognition performance than using Dempster Shafer evidence theory. The reason is that in using Dempster Shafer evidence theory it is implicitly assumed that the component classifiers are independent. Independence is just approximation to reality. If the approximation is too rough, poor results will be obtained. There is no such assumption in the dependent evidence model, so it can efficiently utilize dependence of component classifiers and give a better combined recognition performance.