Genetic Algorithms for MLP Neural Network parameters optimization

In this paper, a hybrid learning algorithm for a multilayer perceptrons (MLP) neural network using genetic algorithms (GA) is proposed. This hybrid learning algorithm has two steps: First, all the parameters (weights and biases) of the initial neural network are encoded to form a long chromosome and tuned by the GA. Second, as a result of the GA process, a quasi-Newton method called Broyden-Fletcher-Goldfarb-Shannon (BFGS) method is applied to train the neural network. Simulation studies on function approximation and nonlinear dynamic system identification are presented to illustrate the performance of the proposed learning algorithm.