Modelling fuzzy production rules with fuzzy expert networks

摘要

The strength of a fuzzy expert system comes from its ability to handle imprecise, uncertain and vague information used by human experts while the power of neural networks lies in their learning, generalization and fault tolerance capabilities. There have already been many attempts to model and formulate fuzzy production rules (FPRs) by using neural networks so that a new system, called a hybrid system, can be developed and will have the advantages of both. The modelling or formulating process, however, is not an easy task. There are many problems that need to be resolved before such a hybrid system can achieve its goal of having the power of both systems. These problems include how to model FPR using a neural network, what necessary modifications to the learning algorithm of the neural network need to be done if the neural network is to have the same inference mechanism as that of a fuzzy expert system, and where could such a hybrid system be applied. In this paper, the necessary network structure, forward and backward processes of a fuzzy expert network (FEN) used to solve these problems will be presented. This FEN had been proposed in Tsang and Yeung (1996, World Congress on Neural Networks, pp. 500–503) but whose details in terms of network structure, forward and backward reasoning mechanism have not been covered. Therefore, this paper aims to introduce the concept of how FEN can formulate and model FPRs. One of its applications is to help knowledge engineers fine-tune knowledge representation parameters such as certainty factor of a rule, threshold value of a proposition and membership values of a fuzzy set. An experiment will also be performed to demonstrate the tuning capability of this FEN.