IMPROVED DESIGN OF NONLINEAR CONTROLLERS USING RECURRENT NEURAL NETWORKS.

Abstract

Recently, there was a growing interest in the field of nonlinear control using neural networks. It has been shown by several researchers that neural networks controllers provide several advantages compared to traditional controllers. In this thesis, the different current available neural network architectures, the advantages and disadvantages of each architecture and its corresponding applications were reviewed. Also, the theory and practice of applying neural networks in the identification and control of nonlinear dynamical systems were discussed. The different cases of dynamical systems and their corresponding methods of identification and control using neural networks were reviewed. Then, a proposal for a new recurrent neural network controller architecture for on-line applications was introduced. This proposed neural network controller architecture is based on the adaptation of the sigmoid weight function of the hidden layer neurons. The structure of the hidden layer neurons of this new proposed architecture and its dynamics were presented. Also, a new dynamic 4>. backpropagation training algorithm to train the different weights of this new proposed architecture including the additional sigmoid function weight was developed. Simulation studies carried out using both the standard neural network controller architecture and the new proposed neural network controller architecture showed that the accuracy and speed of convergence have been improved while using the proposed neural network controller architecture. Also, the number of neurons in the hidden layer of the neural network controller can be minimized using this new proposed architecture.