On Exponentially Weighted Recursive Least Squares for.pdfVIP

On Exponentially Weighted Recursive Least Squares for Estimating Time-Varying Parameters Ta-Hsin Li Department of Mathematical Sciences IBM T. J. Watson Research Center Yorktown Heights, NY 10598-0218  December 9, 2003 Abstract The exponentially weighted recursive least-squares (RLS) has a long history as an algorithm to track time- varying parameters in signal processing and time series analysis. By reviewing the optimality conditions of RLS under a regression framework, possible sources of suboptimality of RLS for tracking time-varying parameters, especially when the parameters satisfy a state-space model, are identified. A straightforward relationship between the RLS variables and the Kalman filtering variables is established under the state- space model assumption. This relationship enables a unified development of several simple algorithms that generalize and extend the traditional RLS. Numerical examples are given to demonstrate the improved tracking performance of theses algorithms. I. INTRODUCTION The method of exponentially weighted recursive least-squares, or simply RLS, has long been employed as a simple alternative to Kalman filtering (KF) for tracking time-varying parameters (e.g., [1]–[6]). KF as an estimator is well known to be optimal under the state-space model (SSM) assumption ([7], [8]). One of the shortcomings of KF is the requirement of complete prior knowledge of the SSM and its parameters. Although under certain conditions the SSM parameters can be estimated from a set of observations via, for example, the Gaussian maximum like