Identification and control of dynamic systems using neural networks.

The aim of this thesis is to contribute in solving problems related to the on-line identification and control of unknown dynamic systems using feedforward neural networks. In this sense, this thesis presents new on-line learning algorithms for feedforward neural networks based upon the theory of variable structure system design, along with mathematical proofs regarding the convergence of solutions given by the algorithms; the boundedness of these solutions; and robustness features of the algorithms with respect to external perturbations affecting the neural networks' signals. In the thesis, the problems of on-line identification of the forward transfer operator, and the inverse transfer operator of unknown dynamic systems are also analysed, and neural networks-based identification schemes are proposed. These identification schemes are tested by computer simulations on linear and nonlinear unknown plants using both continuous-time and discrete-time versions of the proposed learning algorithms. The thesis reports about the direct inverse dynamics control problems using neural networks, and contributes towards solving these problems by proposing a direct inverse dynamics neural network-based control scheme with on-line learning capabilities of the inverse dynamics of the plant, and the addition of a feedback path that enables the resulting control scheme to exhibit robustness characteristics with respect to external disturbances affecting the output of the system. Computer simulation results on the performance of the mentioned control scheme in controlling linear and nonlinear plants are also included. The thesis also formulates a neural network-based internal model control scheme with on-line estimation capabilities of the forward transfer operator and the inverse transfer operator of unknown dynamic systems. The performance of this internal model control scheme is tested by computer simulations using a stable open-loop unknown plant with output signal corrupted by white noise. Finally, the thesis proposes a neural network-based adaptive control scheme where identification and control are simultaneously carried out