Title:The Weighted Null-Space Fitting Method

Abstract:In system identification of structured models, the prediction error method provides asymptotically efficient estimates under mild assumptions, but in general it requires solving a non-convex optimization problem. An alternative class of methods uses a non-parametric model as intermediate step to obtain the model of interest. The weighted null-space fitting (WNSF) method belongs to this class. It is a weighted least-squares method consisting of the following three steps. In the first step, a high-order ARX model is estimated. In the second step, this high-order estimate is reduced to a parametric estimate, with least squares. In the third step, the parametric model is re-estimated, with weighted least squares. The method is flexible in parametrization and suitable for both open- and closed-loop data. In this paper, we show that WNSF provides consistent and asymptotically efficient estimates. Also, simulation studies indicate that WNSF may be competitive with state-of-the-art methods.