NSGA2PI, the hybrid algorithm for Radial Basis function networks

Abstract

This paper presents a multi-objectives optimization algorithm called NSGA2PI for designing the Radial Basis Function Networks (RBFNs) based on the non-dominated sorting genetic algorithms (NSGA-II) and pseudo inverse method. The main considered objectives are: higher classification ability and simpler structure network. NSGA2PI adjusts the RBF layer parameters by using an improved version of NSGA-II and adapts the weights of the output layer by the pseudo inverse method. NSGA2PI is tested on four of the best-known and most widely used data sets, from the University of California at Irvine. We report the values of mean square error, number of hidden nodes, accuracy, sensitivity and specificity. The obtained results are compared against the some best methods which are proposed in the literature. The experiments show that the proposed method obtains RBFNs with higher classification performance, and the more simple structures in comparison to the other methods.