Improving the generalization ability of neural networks by interval arithmetic

Abstract

Recently, interval-arithmetic-based neural networks have been proposed for handling intervals as inputs of multilayer feedforward neural networks. This paper demonstrates that interval arithmetic can be utilized for improving the generalization ability of neural networks for pattern classification problems. We examine two approaches, each of which is used in the classification phase of new patterns and in the learning phase of neural networks, respectively. In the first approach, an interval input vector is generated from a new pattern by adding a certain width to its attribute values. In the second approach, neural networks are trained by interval input vectors generated from training patterns. These approaches are illustrated by a two-dimensional pattern classification problem. The effectiveness of these approaches is examined by computer simulations on a commonly used benchmark data set.