Model-driven online parameter adjustment for zero-attracting LMS

• Published in: Signal processing Vol. 152; pp. 373 - 383
• Main Authors: Jin, Danqi, Chen, Jie, Richard, Cédric, Chen, Jingdong
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.11.2018; Elsevier
• Subjects: Adaptive algorithms; Complex-valued signal; Engineering Sciences; Machine Learning; RZA-LMS; Signal and Image processing; Sparse system identification; Statistics; Transient behavior model; Variable parameter strategy; ZA-LMS; ZA-LMS; RZA-LMS; Transient behavior model; Adaptive algorithms; Sparse system identification; Variable parameter strategy; Complex-valued signal
• ISSN: 0165-1684; 1872-7557
• DOI: 10.1016/j.sigpro.2018.06.020

•This work is the first one in the literature that derives a variable parameter strategy based on a model of the filter performance.•This work provides an online adaptive parameter adjustment strategy for ZA-LMS/RZA-LMS.•The proposed framework can be extended without much difficulty to several other adaptive filters having similar structure. Zero-attracting least-mean-square (ZA-LMS) algorithm has been widely used for online sparse system identification. Similarly to most adaptive filtering algorithms and sparsity-inducing regularization techniques, ZA-LMS appears to face a trade-off between convergence speed and steady-state performance, and between sparsity level and estimation bias. It is therefore important, but not trivial, to optimally set the algorithm parameters. To address this issue, a variable-parameter ZA-LMS algorithm is proposed in this paper, based on a model of the stochastic transient behavior of the ZA-LMS. By minimizing the excess mean-square error (EMSE) at each iteration on the basis of a white input assumption, we obtain closed-form expression of the step-size and regularization parameter. To improve the performance, we introduce the same strategy for the reweighted ZA-LMS (RZA-LMS). Simulation results illustrate the effectiveness of the proposed algorithms and highlight their performance through comparisons with state-of-the-art algorithms, in the case of white and correlated inputs.