A Real-Time Performance Evaluation Indicators of LQG Controller for Systems With Unknown Changes

Control performance is the reference for judging whether the current control strategy suits the dynamic system affected by unknown causes of changes. This article proposes an online indicator of the linear quadratic regulator with Gaussian noise (LQG) to deal with this kind of problem. Due to the poor robustness of LQG, unknown changes, such as model parameter mismatch, non-Gaussian noise, and shifts in both noise mean and covariance, will lead to the current control input deviating from the optimal one, thus causing the degradation of control performance. Therefore, a real-time indicator is urgently needed to calculate the degree of deviation of control performance. In particular, the error between the current control input and the optimal one is approximated by a high-order differentiation. Then, we obtain the indicator by establishing an analytic relation between the model changes and the shift in control performance. The main advantage of the proposed indicator is its ability to provide quantified evaluation results of control performance in real time without requiring information on various changes. A experiment is provided to demonstrate the effectiveness of the proposed evaluation method.