Decision Rule Extraction from Trained Neural Networks Using Rough Sets

The ability of artificial neural networks to learn and generalize complex relationships from a collection of training examples has been established through numerous research studies in recent years. The knowledge acquired by neural networks, however, is considered incomprehensible and not transferable to other knowledge representation schemes such as expert or rule-based systems. Furthermore, the incomprehensibility of knowledge acquired by a neural network prevents users to gain better understanding of a classification task learned by the network. The aim of the present paper is to describe a method that can help to make the knowledge embedded in a trained neural network comprehensible, and thus transform neural networks into a powerful knowledge acquisition tool. Our method is based on rough sets, which offer a useful framework to reason about classification knowledge but lack in generalization capabilities. Unlike many existing methods that require training examples as well as the trained network to extract the knowledge embedded in numerical weights, our method works only with the weight matrix of trained network. No training examples are required. The suggested method has been applied to several trained neural networks with great success.