Interval Arithmetic Perceptron with pruning capability

An Interval Arithmetic Perceptron (IAP) is a neural network where the usual perceptron topology uses the intervals as numerical representation of internal parameters and output: in this paper we present the learning of such a network and show its performance in weight pruning and in the consequent identification of the importance of input features. The applications to the Iris and the Breast Cancer database are presented and simple and efficient classification rules are obtained. IAP compares favourably with a well-established feature screening method.