A Research On Learning Weights Of Fuzzy Production Rules Based On Maximum Fuzzy Entropy

Fuzzy Production Rules (FPRs) are a fundamental and important way of imprecise knowledge representation. For enhancing generalization capability of FPRs for the given examples, the concept of weight is introduced into FPRs, It is necessary to explore specific criterion for determining these weight values. Generally speaking, the usual criterion of the weight values adjustment, which is based only on improving training accuracy, often results in an over-fitting. This paper aims to accomplish this task by using a new method based on the well-known Maximum Fuzzy Entropy Principle. In the case that the training accuracy does not decrease, the testing accuracy will increase with the value of fuzzy entropy of training set. At the same time, adjusting the weight values can change the fuzzy entropy of training set. According to MFEP, learning these parameters can be realized by solving a constrained optimization problem, and then Genetic Algorithm (GA) may be used to determine these parameters. The experimental results show that this new criterion can decrease the phenomenon of over-fitting and can improve the testing accuracy.