93 , p . p . 2706 - 2709 , Oct . 1993 , Nagoya , Japan . 1 A GENETIC ALGORITHM FOR TRAINING RECURRENTNEURAL NETWORKS

A hybrid genetic algorithm is proposed for training neural networks with recurrent connections. A fully connected recurrent ANN model is employed and tested over a number of various problems. Simulation results are presented for three problems: generation of a stable limit cycle, sequence recognition and storage and reproduction of temporal sequences. 1.Introduction Although the recurrent ANN models, seem to be promising, in solving problems associated with time, they suffer from lack of efficient training algorithms. A number of algorithms have been proposed in the past [1-4], for different models of ANNs with recurrent connections. The proposed training algorithms seem to have a limited scope. In this paper we present a hybrid genetic algorithm for training ANNs, which is robust and exhibits enhanced training abilities in a range of difficult problems. 2.Network model We assume a fully connected recurrent neural network that consists of sigmoid units. Let W denote the weight matrix of the network. The topology of the ANN is shown in figure 1. The dimension of weight matrix W is n x (n+m+1), where n is the total number of units and m is the number of input lines (the neuron thresholds are trainable). The total number of weights is N = n.(n+m+1). If yj(t)is the output of jth unit at time t and xi(t)is the value of ith input line at the same time, then the total input to the kth unit at time t is given by the