Classification neural networks improvement using genetic algorithms

The aim of the present work consists the development of a procedure capable to realize a high accurate classification neural network with a reduced number of neurons. In order to avoid local minima, the weights of a proposed three-layer network were computed using a common genetic algorithm. The performances of the genetic algorithm were strongly improved by several means that make an increased balance between exploration and exploitation of the search space. Thus, in order to decrease the number of genes, respectively weights, the dimension of the output vector was minimized by identification of classes with binary numbers. A very favorable effect was obtained by seeding the initial population with a good chromosome obtained by the use of the classical delta-rule learning procedure. It was also investigated the effect of the initial population size, the bounds imposed to the weights of the inter-neuronal connections, the number of neurons in the hidden layer, fitness expression, etc.