Adaptive Regularization Matrix for Affine Projection Algorithm

Abstract

We propose an adaptive regularized affine projection algorithm (AR-APA) with a regularization matrix. Instead of the conventional fixed and weighted identity matrix for regularization, the proposed AR-APA incorporates a nonidentity regularization matrix which is dynamically updated. The adaptation of the regularization matrix is accomplished by the normalized stochastic gradient of mean-square error. Consequently, an efficient and robust APA is derived. Experimental results comparing the proposed AR-APA to the conventional algorithms clearly indicate its superior performance.