Treinamento de redes neurais artificiais baseado em sistemas de estrutura variável com taxa de aprendizado adaptativa

Abstract

This work presents new algorithms for training multilayer perceptron artificial neural networks based on stability properties of sliding mode variable structure systems. The main feature of the proposed algorithms is the adaptability of the gain (learning rate), which is obtained from each update step of the network weights, without the use of heuristics methods to obtain this gain. Two algorithms for continuous time learning multilayer perceptron artificial neural networks with two layer and with linear output layer are developed, allowing the neural network continuously to adapt the network parameters following the input signal variation. The proposed algorithms pursue the same methodology to obtain the adaptive gain. The difierences between them are related with the sliding mode definition and the network weight update rule. In such a manner, the first algorithm is associated with multiple output networks, and the second is used only with the single output networks. In its turn, the second algorithm update the network weights using one expression that guarantee the asymptotical stability around the global minimum weight according to the Lyapunov stability theory. In order to verify the performance of the proposed algorithms, both algorithms were applied to periodic function approximation and induction motor drive. In this last application, the neural network was used as neurocontroller and as induction motor stator flux neural observer. These applications need that neural training has to be made in continuous time, imposing a continuous network weight update according to the overall system requirements. Therefore, the algorithms present two interesting features: easy to use, without the necessity to choose the learning rate parameter by designer; and, adaptive behaviour, without requiring any information about mathematical model of the overall system.