A Reinforcement Learning based Adaptive Supervisor for Multiple Model Adaptive Estimation and Control

Multiple Model Adaptive Estimation and Control (MMAE) has proven to be an effective tool for fault detection and handling. Through the modeling of expected fault dynamics and the design of associated controllers, quick adaptation and significant improvements in system reliability can be realized. However, this technique is hindered by its lack of high level adaptability and considerable implementation costs. In this paper, a reinforcement learning based MMAE supervisor is proposed which allows for increased design flexibility and continual improvement. The ability of the adaptive supervisor to perform fault handling is demonstrated using pitot tube and accelerometer failures in simulation on the Flying Fish autonomous unmanned sea plane. Results show that the adaptive supervisor achieves a 26% increase in normalized aircraft survival probability compared to the nominal controller.