CHANNEL ESTIMATION USING GAUSSIAN PROCESS REGRESSION

Abstract

Gaussian process (GP) regression can be used in the interpolation of observed periodic channel estimates in OFDM transmission systems over both time and frequency in small-scale fading environments. Previous GP regression studies used the popular radial basis function as the GP kernel. In this study, we examine the performance of GP regression using a Bessel kernel with a semi-static hyperparameter vector. Results show that GP regression using the Bessel kernel outperforms the radial basis kernel, as well as traditional interpolation methods such as cubic spline and FIR interpolation, especially when training symbols are spaced far apart in time with respect to the channel coherence time.