Treinamento de redes neurais artificiais utilizando algoritmos genéticos em plataforma distribuída

Abstract

Artificial Neural Networks are algorithms inspired in the functioning of the nervous system that have many applications such as data classification and time series prediction. Its use depends on a training process which, in some problems, can consume significant processing time. This study aimed to evaluate the use of a parallel model of Genetic Algorithm to execute the training of Multilayer Perceptron networks. The analysis was performed by two steps: evaluation of two types of genetic algorithms by comparison among results obtained by Backpropagation learning algorithm; and evaluation of the distributed system by tests running on up to five computers. For the tests, six databases from UCI machine learning repository were used. The Genetic Algorithm has underperformed Backpropagation; however it was able to find solutions with good generalization capacity. The total average training time was reduced to 31% when the distributed system was applied. Despite the low gain in performance, the parallel model increases the stability of the evolutionary learning process.