The Generalization Ability Of Weighted Fuzzy If-then Rules Based On Fuzzy Entropy Maximum

The generalization capability (GC) of an expert inductive system, which is regarded as the most important issue for evaluating the performance of the expert inductive system, is defined as the classifying ability of the system to these excluded objects from training process. The GC of the expert inductive system depends mainly on the expert knowledge, which are usually represented as fuzzy production rules (FPRs), inside the expert inductive system. As a consequent, the acquiring and refining of the FPRs are the key for us to improve the performance of the inductive system. It has been realized that it is an arduous task for us to improve the GC of the expert inductive systems in artificial intelligence area. The most existing approaches to rule refinement are based on the further reduction of training error, which is really helpful to improve the training accuracy. But it is very likely to lead to an over-fitting and therefore seriously downgrades the GC of the FPRs. In order to improve the capability of knowledge representation of FPRs, several parameters such as local weight, global weight and certainty factor have been incorporated into the FPRs. Regarding the weights and certainty factors as adjustable parameters, firstly, this paper explores the relationship between these adjustable parameters and the GC. Secondly, this paper proposes a new rule refinement scheme based on the well known fuzzy entropy maximization. Experimental results on a number of selected databases demonstrate the expected improvement of GA of the FPR-based expert systems.