This project will attempt to develop new concepts, methods, and algorithms for the servomechanism of nonlinear dynamic systems using artificial neural networks. The main thrust of the research is directed toward the investigation of incorporating neural network controllers with nonlinear plants whose dynamics are unknown. These controllers can be implemented by the modern systolic array and parallel processing architectures. Servo compensator design for a general class of nonlinear stable plants is introduced. Servo compensators in general require the knowledge of the input/output Jacobian of the plant. An off-line trained neural network can be made to implement the steady-state input/output Jacobian of the nonlinear plant. The new servo scheme, utilizing the concepts of integral manifold and singular perturbation, achieves constant set-point tracking in plants with unknown dynamics. For tracking dynamic signals and robustness improvements, on-line specialized training is introduced. Decentralized experiments will be developed for estimating the input/output characteristics of multi-input-multi-output unknown nonlinear plants using neural networks for subsequent decentralized servo objectives. This is an extension of the linear scheme that is valid only in unknown linear systems.
