Extracting rules from a GA-pruned neural network

Although artificial neural networks have been proven to be a powerful and general technique, they are often regarded as "black box" and the result of a network may not be easily applied to related problems. In this paper, a genetic algorithm is used to prune a trained network, then the pruned network is converted to M trees where M is the number of the output units and equal to the number of classes of the problem. Finally, rule sets are extracted for each class by analyzing each tree. In this way, we managed to utilize the advantages of artificial neural networks, genetic algorithms and symbolic learning, and avoid some of their disadvantages at the same time. The rules are shown to be interpretable while preserving the accuracy of the networks and can be easily used in related fields.