An IV-SVM-based approach for identification of state-space LPV models under generic noise conditions

This paper presents a nonparametric identification method for state-space linear parameter-varying (LPV) models using a modified support vector machine (SVM) approach. While most LPV identification schemes in the state-space form fall under the general category of parametric methods, regularization-based SVMs provide a viable alternative to model scheduling dependencies, without the need of specifying the dependency structure and with an attractive bias-variance trade-off. In this paper, a solution is proposed for nonparametric identification of LPV state-space models in terms of least-squares SVMs (LS-SVM) and is then extended in a way that the proposed estimation is robust to errors in the noise model estimation. The so-called instrumental variables (IV) method has been used in linear system identification for quite some time, and has recently seen its application in the identification of both nonlinear and LPV systems in the input-output (IO) form. The IV method reduces the bias in estimated LPV state-space models in case the noise model is not estimated properly or is unknown. In the proposed method of this paper, the attractive bias-variance trade-off properties of LS-SVMs are combined with statistical properties of IV-based methods to give robust estimates of the functional dependencies. Numerical examples are provided to compare the performances of the proposed IV-based technique with the LS-SVM-based LPV model identification methods.