GpGp: GpGp: Fast Gaussian Process Computing.

In GpGp: Fast Gaussian Process Computation Using Vecchia's Approximation

GpGp: Fast Gaussian Process Computing.

Description

Vecchia's (1988) Gaussian process approximation has emerged among its competitors as a leader in computational scalability and accuracy. This package includes implementations of the original approximation, as well as several updates to it, including the reordered and grouped versions of the approximation outlined in Guinness (2018) and the Fisher scoring algorithm described in Guinness (2019). The package supports spatial models, spatial-temporal models, models on spheres, and some nonstationary models.

Details

The main functions of the package are fit_model, and predictions. fit_model returns estimates of covariance parameters and linear mean parameters. The user is expected to select a covariance function and specify it with a string. Currently supported covariance functions are

• matern_isotropic

• exponential_isotropic

• matern_anisotropic2D

• exponential_anisotropic2D

• matern_anisotropic3D

• exponential_anisotropic3D

• matern_anisotropic3D_alt

• matern15_isotropic

• matern25_isotropic

• matern35_isotropic

• matern45_isotropic

• matern_scaledim

• exponential_scaledim

• matern15_scaledim

• matern25_scaledim

• matern35_scaledim

• matern45_scaledim

• matern_spacetime

• exponential_spacetime

• matern_nonstat_var

• exponential_nonstat_var

• matern_sphere

• exponential_sphere

• matern_spheretime

• exponential_spheretime

• matern_sphere_warp

• exponential_sphere_warp

• matern_spheretime_warp

• exponential_spheretime_warp

If there are covariates, they can be expressed via a design matrix X, each row containing the covariates corresponding to the same row in locs.

For predictions, the user should specify prediction locations locs_pred and a prediction design matrix X_pred.

The vignettes are intended to be helpful for getting a sense of how these functions work.

For Gaussian process researchers, the package also provides access to functions for computing the likelihood, gradient, and Fisher information with respect to covariance parameters; reordering functions, nearest neighbor-finding functions, grouping (partitioning) functions, and approximate simulation functions.

Author(s)

Maintainer: Joseph Guinness [e-mail]

Authors:

• Matthias Katzfuss [e-mail]

• Youssef Fahmy [e-mail]

GpGp documentation built on Oct. 16, 2024, 9:06 a.m.