Evolving Generalized Euclidean Distances for Training RBNN

Abstract

In Radial Basis Neural Networks (RBNN), the activation of each neuron depends on the Euclidean distance between a pattern and the neuron center. Such a symmetrical activation assumes that all attributes are equally relevant, which might not be true. Non-symmetrical distances like Mahalanobis can be used. However, this distance is computed directly from the data covariance matrix and therefore the accuracy of the learning algorithm is not taken into account. In this paper, we propose to use a Genetic Algorithm to search for a generalized Euclidean distance matrix, that minimizes the error produced by a RBNN.