A neural network system for adaptive control

This dissertation considers the problem of controlling linear and nonlinear dynamic systems subject to parameter variations and disturbances with the use of artificial neural networks. In general, the implementation of complex controllers for such systems has required information concerning plant parameters and their dynamics that was acquired either prior to implementation, or while the controller is on-line. Conventional adaptive controller methodology is largely limited to linear or linearizable systems, while applications to nonlinear systems have introduced a significant increase in complexity, so that such controllers may not be amenable at all for nonlinear systems. This dissertation presents a novel controller design approach based on an Artificial Neural Network System (ANNS), a combination of neural networks, using a new architecture that effectively and adaptively controls complex linear and nonlinear multiple-input-multiple-output (MIMO) systems in the presence of parameter variations and disturbances. The ANNS controller first uses an off-line training regimen based on the traditional back propagation (BP) weight-training procedure, and then uses on-line training to train a set of neural network controllers where each is applicable to a particular range of the dynamic variables of the system. Finally, the ANNS intelligently selects, on-line, the appropriate neural network controller to generate the control signal at any given time. Simulation results for both linear and nonlinear systems are presented to demonstrate the effectiveness of this approach.