ECS-9812213 Krogh In this research project a new class of switching controllers will be investigated. Given multiple controllers for a plant, each designed for possibly different operating regions with distinct regions of stability and performance characteristics, the controller to be applied at each sampling instant will be selected according to on-line evaluations of the future potential closed-loop system performance for each of the controllers. The switching controller will achieve a larger operating region and better performance than could be realized using any one of the controllers alone. Neural networks will be trained to estimate cost-to-go functions and stability regions based on experimental data using techniques from neural dynamic programming. State feedback and output feedback controllers will be studied as will as algorithms for on-line learning of performance measures. Policy iteration techniques will be investigated for optimizing the switching strategies. Closed-loop stability and performance will be studied drawing on the theory of stability and invariant sets in hybrid dynamic systems. The switching control strategies will be demonstrated using simulation.
