In recent years, the evidence reasoning theory proposed by Dempster in 1967, has made great progress. However, there are still some problems in theory and applications. This thesis focuses on the following fields such as the classification of evidence combination rules, evaluation criteria, etc. The main contributions are as follows:1. We summarize the latest theory and application works in this field. Based on analyzing evidence combination rules in detail, they are classified as four classes, which are model improvement method, conflict-information-distribution method, all-information-distribution method and evidences improvement method. The research shows that it is more reasonable and feasible than other classification methods.2. Based on the concepts of main focus and polarization range, the polarization of evidence combination is proposed for evaluating the capabilities of fusing low-conflict evidences in evidence combination. Anti-high-conflict and complexity are proposed for evaluating the capabilities of fusing high-conflict evidences and computational complexity in evidence combination. The simulations and analyses prove that these evaluation criteria can distinguish the capabilities of evidence combination rules, and supply the foundation for constructing and applying these combination rules.3. The evaluation criteria of identity and focus are perfected. Based on the basic properties and four proposed properties of evidence combination, the evaluation system of evidence combination is founded including 3 parts and 9 evaluation criteria. It can be used to evaluate the validity of evidence combination rules.4. According to proposed evaluation system, twelve typical methods of evidence combination are analyzed with many typical examples. The conclusions show that the evaluation system and typical examples can be use to reasonably evaluate evidence combination rules from different degree5. We analyze and summarize the problems of constructing basic belief assignment and decision methods in application of evidence reasoning. The recognition of sequential images is done based on BP neural network and evidence combination rules. The results show that the polarization has important significancefor selecting reasonable evidence combination rules. |