TRAINING MLP NEURAL NETWORKS BY MEANS OF GENETIC ALGORITHMS

摘要

Despite the success of neural networks, their design is still largely performed through a consequent process of definition of the architecture and a training process to adjust the weights. Recently, new methods based on evolutionary computation have been applied to the synthesis of neural networks. Genetic algorithms are powerful search techniques that can by applied for designing neural networks, by replacing traditional training algorithms. This work starts from the Neural Network based System Identification toolbox and extends it with a technique based on genetic algorithms. GAs encode the network as individuals in a two-dimensional way, one dimension coding the weights and the other the type of the activation function. Thus, GAs are used for adjusting the weights and simultaneous search for the activation function of each network's node. More experiments concerning network training by means of GAs are performed and some experimental results are presented.