Abstract:
Artificial neural networks opened a new horizon in many research areas. This understanding, also, brought a new way of thinking into the concept of control. The ever-increasing technological demands steered the control engineers to design more sophisticated controllers. In this respect, artificial neural networks were proposed as a new approach because of their massively parallel data processing properties, adaptiveness and powerful mapping capabilities. Especially the learning property of these networks made them extremely attractive. There are various methods that are used for the training of artificial neural networks. Two of them are included in this study. These are, namely, the backpropagation method and the Levenberg-Marquardt optimization technique. The learning time for the former is excessively long especially around the minima since it uses the first order derivatives of the performance function, while in the latter the learning time is very short because of the extra information coming form the second derivatives of the performance function. The computational complexity and the hardware requirements are large for the latter. The identification of nonlinear dynamical systems is a substantial part of the controller training therefore it is included in this work and discussed in the simulation results. The main idea that lies under the procedure is obtaining a regular and mathematically tractable model of the system which is of interest.Based on these two learning methods and system identification, controller design methodology for error backpropagation, inverse control, self-tuning control, model reference adaptive control, self-learning control and dynamical neural unit based control methodologies are explained and excessive simulation results are discussed in this thesis.
