Comparison of neural and control theoretic techniques for nonlinear dynamic systems

Abstract

This thesis compares classical nonlinear control theoretic techniques with recently developed neural network control methods based on the simulation and experimental results on a simple electromechanical system. The system has a configuration-dependent inertia, which contributes a substantial nonlinearity. The controllers being studied include PID, sliding control, adaptive sliding control, and two different controllers based on neural networks: one uses feedback error learning approach while the other uses a Gaussian network control method. The Gaussian network controller is tested only in simulation due to lack of time. These controllers are evaluated based on the amount of a priori knowledge required, tracking performance, stability guarantees, and computational requirements. Suggestions for choosing appropriate control techniques to one's specific control applications are provided based on these partial comparison results.