Kernel least mean square algorithm with mixed kernel

Abstract

This paper presents a novel kernel least square algorithm with mixed kernel (KLMS-MK) to improve the filtering performance of kernel least mean square (KLMS). By applying the convex combination method to the kernel function in KLMS, KLMS-MK bears the advantages of both the Gaussian kernel and the Laplace kernel. In KLMS-MK, the mixed parameter for the convex combination is updated with the stochastic gradient descent. Therefore, the steady-state mean square error (MSE) and the convergence rate are improved by KLMS-MK, simultaneously. Simulation results on chaotic time series prediction and nonlinear regression validate the excellent performance of KLMS-MK from the aspects of the convergence rate and estimation accuracy.