New method of RKHS model complexity reduction

Abstract

In this paper, we propose a new kernel method with complexity reduction of reproducing kernel Hilbert space (RKHS) models. In RKHS models, the number of parameters is equal to the training set size; this leads to a complex representation. We propose a new method, the reduced kernel canonical correlation analysis (RKCCA), to reduce the number of parameters of RKHS models. This method consists on approximating the canonical correlation coefficients by a set of observation data. The proposed method is used to identify experimentally two nonlinear systems.