In Proceedings of STeP ' 96 { Genes , Nets and Symbols

A main criterion for the accuracy of solutions of an Artiicial Neural Network (ANN) for classiica-tion tasks is the architecture. In order to nd problem-adapted topologies of ANNs, we adopted the evolutionary approach to ANN design by employing a Genetic Algorithm (GA) to evolve ANNs which are represented using a direct encoding method. As ANNs of low complexity show better generalization capabilities than more complex networks, we incorporated a regularization term into the tness function which together with the problem representation is determining the preferred regions of the search space the GA will focus on. Especially, we investigated two diierent regularization terms proposed in literature and experimented with various degrees of impact on the tness function. The parallel net-GEN system which has been implemented by the authors is generating problem{adapted Feed{Forward ANNs being trained by Error{Back{Propagation. Empirical results on a real world problem taken from the PROBEN1 ANN benchmark suite are presented.