Title: Score function approximation for Gaussian processes using equivalent kernels
Authors: William Kleiber - University of Colorado (United States)
Zachary Mullen - University of Colorado (United States) [presenting]
Abstract: Large, spatially indexed datasets bring with them a host of computational and mathematical challenges. Parameter estimation often relies on maximum likelihood, which, for Gaussian processes, involves matrix manipulations of the covariance matrix including solving systems of equations and determinant calculations. The score function, on the other hand, avoids direct calculation of the determinant, but still requires solving a large number of linear equations. We propose an equivalent kernel approximation to the score function of a Gaussian process. A nugget effect is required for the approximation, and we find the nugget must not be negligible for adequate approximations. For large, noisy processes, the approximation is fast, accurate, and compares well against existing approaches.