Switching controllers based on neural network estimates of stability regions and controller performance

This paper presents new results on switching control using neural networks. Given a set of candidate controllers, a pair of neural networks is trained to identify the stability region and estimate the closed-loop performance for each controller. The neural network outputs are used in the on-line switching rule to select the controller output to be applied to the system during each control period. The paper presents architectures and training procedures for the neural networks and sufficient conditions for stability of the closed-loop system using the proposed switching strategy. The neural-network-based switching strategy is applied to generate the switching strategy embeded in the SIMPLEX architecture, a real-time infrastructure for soft on-line control system upgrades. Results are shown for the real-time level control of a submerged vessel.