The partitioned kernel machine algorithm for online learning

Abstract

Kernel machines have been successfully applied to many engineering problems requiring pattern recognition and regression. Kernel machines are a family of machine learning algorithms including support vector machines (SVM) [1], kernel least mean squares adaptive filter (KLMS) [2], and kernel recursive least squares (KRLS) adaptive filter [3] to name a few. In this paper we present the partitioned kernel machine algorithm for use in online learning in virtual environments. The PKM algorithm enhances the accuracy of the computationally efficient KLMS algorithm. The PKM algorithm is an iterative update procedure that focuses on a subset of the stored vectors in the kernel machine buffer. We use a similarity measure for the selection of kernel machine vectors that allow more common vectors to be updated more frequently, and outlier vectors to be updated less frequently. We validate the increased accuracy of our novel algorithm in two separate experimental settings.