An Method to Extract Comprehensible Rules

Knowledge acquisition has been a bottleneck when constructing expert system. Neural network has many advantages to obtain knowledge in the application field. However, the major drawback of neural network is to be lack of comprehensibility. Knowledge obtained by network is concealed in the architecture of neural networks and weights between neurons. Rule extraction from neural network has been recognized one of the proper methods to deal with this drawback. This paper developed researches on rule extraction. For problems with continuous-valued and discrete-valued attributes, the paper presents an approach to extract understandable and concise rules. Rules extracted are comprehensible not only for discrete value but also for continuous value. Our experiment results on real-word dataset validate our approach and show that rules extracted by our approach are comprehensible and concise.