Learning non-linear time scales with Kernel γ-Filters

Abstract

A family of kernel methods, based on the γ-filter structure, is presented for non-linear system identification and time series prediction. The kernel trick allows us to develop the natural non-linear extension of the (linear) support vector machine (SVM) γ-filter, but this approach yields a rigid system model without non-linear cross relation between time-scales. Several functional analysis properties allow us to develop a full, principled family of kernel γ-filters. The improved performance in several application examples suggests that a more appropriate representation of signal states is achieved.