Comparação entre métodos de treinamento de redes neurais artificiais na tarefa de identificação de sistemas

Abstract

In this work, different algorithms for training multilayer perceptron artificial neural networks are compared in the system identification task. To generate the database, a simulation, in Matlab software, of a direct current motor subject to different levels of voltage and load was carried out, and the current and speed of this motor are considered as outputs. Due to its simplicity, the chosen neural network architecture has only one hidden layer, and the comparison was performed between the methods error backpropagation, extreme learning machine in batch and extreme learning machine with online training. All models were implemented in Matlab software and, based on the measurements of mean squared error between real and predicted values, the online training with an extreme learning machine was the most effective in the identification task.