Optimizing RBF networks with cooperative/competitive evolution of units and fuzzy rules

Abstract

This paper presents a new evolutionary method to design optimal networks of Radial Basis Functions (RBFs). The main characteristics of this method lie in the estimation of the fitness of a neurone in the population and in the choice of the operator to apply; for this latter objective, a set of fuzzy rules is used. Thus, the estimation of the fitness considered here, is done by considering three main factors: The weight of the neuron in the RBF Network, the overlapping among neurons, and the distances from neurons to the points where the approximation is worst. These factors allow us to define a fitness function in which concepts such as cooperation, speciation, and niching are taken into account. These three factors are also used as linguistic variables in a fuzzy logic system to choose the operator to apply. The proposed method has been tested with the Mackey-Glass series. © Springer-Verlag Berlin Heidelberg 2001.