PUB - Publikationen an der Universität Bielefeld

Rule extraction from neural networks is a powerful tool for knowledge discovery from data. In order to facilitate rule extraction, trained neural networks are often pruned so that the extracted rules are understandable to human users. This chapter presents a method for extracting interpretable rules from neural networks that are generated using an evolutionary multi-objective algorithm. In the algorithm, the accuracy on the training data and the complexity of the neural networks are minimized simultaneously. Since there is a tradeoff between accuracy and complexity, a number of Pareto-optimal neural networks, instead of one single optimal neural network, are obtained. We show that the Pareto-optimal networks with a minimal degree of complexity are often interpretable in that understandable logic rules can be extracted from them straightforwardly. The proposed approach is verified on two benchmark problems.