A HYBRID LEARNING ALGORITHM

摘要

The integration of symbolic prior knowledge and neural networks in so-called Knowledge and neural networks is becoming increasingly popular for solving difficult real-world problems[1]. Hybrid intelligent systems that combine and artificial neural network systems typically have four phases involving domain knowledge representation, mapping into connectionist network, network training, and rule extraction respectively. In order to obtain a concise set of symbolic rules, redundant and irrelevant units and connections of a trained neural network are usually removed by a network pruning algorithm before rule are extracted Typical pruning algorithms require retraining the network, which incurs additional cost. In this paper, we introduce a new rule extraction technique without network retraining. Our technique is a universal and comprehensive approach that extracts all embedded knowledge in a trained artificial neural network and represents it in a rule base format. Experimental results show that the size and the predictive accuracy of the rule generated are comparable to those extracted by another method, which prunes and retrains the network.