Affine Projection Algorithms with Adaptive Regularization Matrix

We propose a family of novel affine projection algorithms (APA) with adaptive regularization matrix. Conventional regularized APA (R-APA) uses a fixed and weighted identity matrix for regularization. The proposed algorithms incorporate a non-identity regularization matrix which is also dynamically updated. The matrix adaptation is based on the normalized stochastic-gradient of mean-square error. As a result, the efficient and robust algorithms are derived. Throughout experiments, we illustrate that the proposed algorithms show better performance than the conventional R-APA and comparable to the RLS algorithm in terms of the convergence rate and the misadjustment error.