SVR-based Observer Design for Unknown Linear Systems: Complexity and Performance
In this paper we consider estimating the system parameters and designing stable observer for unknown noisy linear time-invariant (LTI) systems. We propose a Support Vector Regression (SVR) based estimator to provide adjustable asymmetric error interval for estimations. This estimator is capable to trade-off bias-variance of the estimation error by tuning parameter $γ> 0$ in the loss function. This method enjoys the same sample complexity of $\mathcal{O}(1/\sqrt{N})$ as the Ordinary Least Square (OLS) based methods but achieves a $\mathcal{O}(1/(γ+1))$ smaller variance. Then, a stable observer gain design procedure based on the estimations is proposed. The observation performance bound based on the estimations is evaluated by the mean square observation error, which is shown to be adjustable by tuning the parameter $γ$, thus achieving higher scalability than the OLS methods. The advantages of the estimation error bias-variance trade-off for observer design are also demonstrated through matrix spectrum and observation performance optimality analysis. Extensive simulation validations are conducted to verify the computed estimation error and performance optimality with different $γ$ and noise settings. The variances of the estimation error and the fluctuations in performance are smaller with a properly-designed parameter $γ$ compared with the OLS methods.