TRAINING RADIAL BASIS FUNCTION NETWORKS BY GENETIC ALGORITHMS

One of the issues of modeling a RBFNN - Radial Basis Function Neural Network consists of deternnning the weights of the output layer, usually represented by a rectangular matrix. The inconvenient characteristic at this stage it's the calculation of the pseudo-inverse of the activation values matrix. This operation may become computationally expensive and cause rounding errors when the amount of variables is large or the activation values form an ill-conditioned matrix so that the model can misclassify the patterns, In our research, Genetic Algorithms for continuous variables determines the weights of the output layer of a RBNN and we've made a comparsion with the traditional method of pseudo-inversion. The proposed approach generates matrices of random normally distributed weights which are individuals of the population and applies the Michalewicz's genetic operators until some stopping criteria is reached. We've tested four classification patterns databases and an overall mean accuracy lies in the range 91-98%. in the best case and 58-63% in the worse case.