Sparsity regularized recursive total least-squares

Abstract

This paper introduces a new family of recursive total least-squares (RTLS) algorithms for identification of sparse systems with noisy input vector. We regularize the RTLS cost function by adding a sparsifying term and utilize subgradient analysis. We present l(1) norm and approximate l(0) norm regularized RTLS algorithms, and we elaborate on the selection of algorithm parameters. Simulation results show that the presented algorithms outperform the existing RLS and RTLS algorithms significantly in terms of mean square deviation (MSD). Furthermore, we demonstrate the virtues of our automatic selection for regularization parameter when l(1) norm regularization is applied. (C) 2015 Elsevier Inc. All rights reserved.