Safe Adaptive Control for Performance Improvement

In the past fifty years, adaptive control theory has reached a high level of maturity. However, applications of adaptive control in safety sensitive systems are still very limited. Conventional identifier based adaptive control may lead to poor transients during the learning process due to the conflict between parameter estimation and control. Their nonlinear nature makes it difficult to check stability and robustness bounds as done in the linear time invariant (LTI) case. As a result in applications where safety is at stake practitioners are reluctant to close the loop using an identifier based adaptive controller. In this paper, we propose a novel adaptive controller, referred to as safe adaptive controller, which is composed of a supervisor and two candidate controllers. One candidate controller is a LTI model reference controller that can always guarantee system stability, while the other is a model reference adaptive controller that can tune its parameters to counteract the changes in the plant. The supervisor evaluates the performance of the two candidate controllers without using any plant model information and activates the candidate controller leading to better transient response. It is shown that the system stability is guaranteed as long as the non-adaptive candidate controller is stabilizing. Simulation results are presented to demonstrate the proposed safe adaptive controller is able to achieve better performance than the two candidate controllers