Online least angle regression algorithm for sparse system identification

Abstract

Traditional least mean square algorithm based system identifications schemes are not effective in identifying sparse systems and a few sparse adaptive algorithms have been developed recently to overcome this limitation. Least angle regression (LARS) algorithm, which has recently emerged as an effective model selection approach, has found applications in sparse coding and sparse signal recovery. LARS has been shown to provide very fast convergence in sparse signal recovery tasks. The offline learning nature of the LARS algorithm prevents the direct use of the algorithm for online sparse system identification. With an objective to utilize the fast convergence offered by LARS algorithm, an online sparse system identification method based on LARS is developed in this letter. The new scheme has been shown to provide improved convergence behaviour over conventional sparse system identification methods.