A comparison between Kriging and radial basis function networks for nonlinear prediction.

Abstract

Predictions by Kriging and radial basis function (RBF) networks with gaussian Kernels are compared. Kriging is a semi--parametric approach that does not rely on any specific model structure, which makes it much more flexible than approaches based on parametric behavioural models. On the other hand, accurate predictions are obtained for short training sequences, which is not the case for nonparametric prediction methods based on neural networks. Examples are presented to illustrate the effectiveness of the method. 1. INTRODUCTION We consider the situation where the relationship beetween the input and output sequences fx k g and fy k g of a SISO system S is dominated by nonlinear characteristics. When using a parametric nonlinear model, first one has to choose a suitable model structure [1]. Second, once a structure has been chosen, one has to estimate its parameters. Traditional parametric representations for nonlinear unknown structures are the Volterra, Wiener or NARMAX (Nonlinear Au...