Identification of Linear Parameter-Varying Models with Unknown Parameter Dependence Using ϵ-Support Vector Regression

The bias-variance trade-off is very sensitive to the prior selection of functional dependencies when identifying linear parameter-varying (LPV) systems. To counteract this difficulty, various non-parametric methods have been recently proposed. These methods avoid the manual selection of the functional dependency but rather learn it during the identification itself. In this paper, we propose an algorithm to implement the ϵ-tube support vector regression (SVR) approach to such an LPV identification problem. We use some results from the machine learning literature to tune the parameters and provide different directions one could take to optimize these parameters. We demonstrate the effectiveness of our method on an example and compare the results with other methods recently proposed. We observe that because of the insensitive ϵ-tube, the number of parameters was greatly reduced still maintaining the same accuracy in terms of the best fit ratio.

Funding(s)

LPV System Identification

Options

Identification of Linear Parameter-Varying Models with Unknown Parameter Dependence Using ϵ-Support Vector Regression