Functions for fitting and doing predictions with
Gaussian process models using Vecchia's (1988) approximation.
Package also includes functions for reordering input locations,
finding ordered nearest neighbors (with help from 'FNN' package),
grouping operations, and conditional simulations.
Covariance functions for spatial and spatial-temporal data
on Euclidean domains and spheres are provided. The original
approximation is due to Vecchia (1988)
< http://www.jstor.org/stable/2345768>, and the reordering and
grouping methods are from Guinness (2018)
GpGp is an R package for fast approximate Gaussian process computation. The package includes implementations of the Vecchia's (1988) original approximation, as well as several updates to it, including the reordered and grouped versions of the approximation outlined in Guinness (2018).
The package can be installed from CRAN with the usual R command
install.packages("GpGp")
or directly from Github for the latest version
devtools::install_github("joeguinness/GpGp")
We always recommend using multithreaded linear algebra libraries in R, but for this package in particular, using multithreaded libraries can have a big impact on performance. On a Mac, there is a very simple way to link to the Apple Accelerate Framework. On PC and Linux, it's more complicated, but you can use Microsoft R Open instead, which comes automatically with multithreaded libraries.
See the vignettes directory for examples using the package. The file vignette_likelihood.R shows how to use the low-level functions to reorder, find neighbors, group, and calculate likelihoods. The file vignette_windspeed.R shows an analysis of spatial-temporal windspeed data using higher-level functions (i.e. more automation).
This is a minor release fixing numerical stability problems that arise during optimization of the likelihood.
Added a check in each of the proflik_mean* functions to avoid inverting the information matrix when it is numerically singular.
Changed the default number of Nelder-Mead iterations in fit_model to 100