Evolução da arquitetura de perceptron multicamadas baseada em otimização por enxame de partículas e colônia artificial de abelhas

Abstract

The present work proposes to develop two hybrid intelligent systems based on Particle Swarm Optimization and Artificial Bee Colony, capable of realizing the optimization of the architecture of artificial neural networks of the multilayer Perceptron type. The internal parameters of the architecture chosen for optimization were those that have the greatest influence on the performance of the network, being: number of hidden layers, number of neurons per layer and learning rate. In this way, to perform the choosing process of the network topology that best fits the set of data to be classified, the bioinspired algorithms use as fitness the average quadratic error generated by the multilayer Perceptron in the test phase. The obtaining and analyze of the proposed systems results were performed in two steps: validation and test, where a total of 192 thousand networks were trained. During the validation stage, the databases of medical diagnoses "Breast Cancer", "Mammographic Mass" and "Thyroid" obtained from the machine learning repository, UCI, were used. While, for the test phase, the databases of electromyographic signals “Classes of Movements” and “Levels of Contraction” were used, treated using the Wavelet Packet Transform of the Haar family. From the obtained results it was possible to verify the efficiency of both systems, which are presented as great tools to assist professionals responsible for the design of artificial neural networks capable of adapting to any class of problems involving the optimization of multilayer Perceptron type network topology.