The decorrelated kernel least-mean-square algorithm

Abstract

With highly correlated input signal, the kernel least-mean-square algorithm(KLMS) always possess a low convergence rate. To overcome this problem the input signal should be decorrelated before adaptive filtering. A decorrelated kernel least-mean-square algorithm(DKLMS) is proposed, which is the combination of KLMS and decorrelation. Using the characteristics of Gaussian kernel, the correlation coefficient is simplified, and the normalized variable step size is obtained and simplified. The iteration process of DKLMS algorithm is presented. The computer simulation results show that DKLMS has a smaller steady-state error and a faster convergence rate than KLMS, and outperforms DLMS.