Performance characterization and controller scheduling for dynamic systems using neural networks

In this thesis we explore the use of neural networks to evaluate the performance of nonlinear systems and to switch among a pool of controllers to achieve a desired closed-loop performance. This method is motivated on the needs of a automated way to generate controller schedules and the development of a control architecture for evolvable systems called SIMPLEX. The method is based on two neural network estimates for each controller, one for its region of stability and another for a performance index. Neurodynamic programming techniques are used for neural network training providing a way to realize the architecture in real-time. Several types of estimates are developed to handle uncertainties and disturbances. Convergence of the algorithms and confidence intervals are analyzed. The architecture and controller scheduling algorithm developed in this thesis allows us to assure stability of the closed-loop system and improve closed-loop performance by combining tested and new controller capabilities. Computational complexity is analyzed. Results are presented in a simulation example and a submerged vessel application. Future lines of research are discussed