A Zero-Attracting Sparse Lncosh Adaptive Algorithm

Performance of adaptive filters highly depends on the eigenvalue spread of the autocorrelation matrix of the input signal. For example, performance of the least-mean-square (LMS) algorithm deteriorates if this eigenvalue spread is relatively high. Recently, a least lncosh (lncosh) algorithm has been proposed to enhance the performance of the LMS algorithm. The algorithm utilizes lncosh function of the error in its cost function. Based on this, we propose a new algorithm for that imposes an l
0
- norm penalty to the cost function of the lncosh algorithm. This penalty term is capable to exploit the system sparsity in system identification settings. The performance of the proposed algorithm has been measured in the presence of correlated and non-correlated input signals. The proposed algorithm has shown significant performance compared to those of the lncosh and re-weighted zero-attracting LMS (RZA-LMS) algorithms in different system identification setting.