Extracting Weighted Fuzzy Production Rules From Trained Neural Networks

It is significant to extract IF-THEN rules from neural networks that have no explicit and declarative knowledge for insight into how decisions are made and knowledge acquisition. This paper proposes a new method to extract weighted fuzzy production rules from trained neural networks for classification problem with numerical condition attributes. In order to translate the knowledge in neural networks into weighted fuzzy production rules, three aspects are considered: (1) designing and training of neural networks; (2) constructing matrix of importance index by analysis of weight, and then building weighted fuzzy production rules; (3) making fuzzy reasoning accorded with weighted fuzzy production rules. A new method is proposed based on the three aspects. The characteristics of the new method are that:(1) building weighted fuzzy production rules and designing and training of neural networks and fuzzy reasoning are considered synthetically; (2) constructing weighted fuzzy production rules based on analysis of weight. The experimental results show that knowledge in neural networks is explained quite accurately by weighted fuzzy production rules and corresponding fuzzy reasoning.