Projected Kernel Recursive Least Squares Algorithm

Abstract

In this paper, a novel sparse kernel recursive least squares algorithm, namely the Projected Kernel Recursive Least Squares (PKRLS) algorithm, is proposed. In PKRLS, a simple online vector projection (VP) method is used to represent the similarity between the current input and the dictionary in a feature space. The use of projection method applies sufficiently the information contained in data to update our solution. Compared with the quantized kernel recursive least squares (QKRLS) algorithm, which is a kind of kernel adaptive filter using vector quantization (VQ) in input space, simulation results validate that PKRLS can achieve a comparable filtering performance in terms of sparse network sizes and testing mean square error.