Robert B. Gramacy Professor of Statistics

Local Approximate Gaussian Process Regression

laGP is an R package providing approximate GP regression for large computer experiments and spatial datasets. The approximation is based on finding small local designs for prediction (independently) at particular inputs.

This software is licensed under the GNU Lesser Public License (LGPL), version 2 or later. See the change log and an archive of previous versions.

The current version provides:

ALC, MSPE and NN-based local approximation, as well as EFI-based global heuristics
local MLE/MAP inference for (isotropic and separable) lengthscales and nuggets
OpenMP for approximation over a vast out-of-sample testing set
GPU acceleration for local ALC subroutine evaluations
SNOW/parallel-package cluster parallelization
computer model calibration via optimization
blackbox constrained optimization via augmented Lagrangians
an interface to lower-level (full) GP inference and prediction

Obtaining the package

Download R from cran.r-project.org by selecting the version for your operating system.
Install the laGP package, from within R.
R> install.packages(c("laGP"))
Optionally, install the mvtnorm, and snow packages, which are helpful for some of the comparisons in the examples and demos.
R> install.packages(c("mvtnorm", "snow"))
Load the library as you would for any R library.
R> library(laGP)

Documentation

The laGP tutorial is implemented as a package vignette, authored in Sweave. The pdf can be obtained from within R with the following code.
R> vignette("laGP")
To obtain the source code contained in the vignette, use the Stangle command.
R> v <- vignette("laGP")
R> Stangle(paste(v$Dir, "/doc/", v$File, sep=""))
The code from Section 4 of the vignette, on Calibration, is available as a standalone demo.
R> demo("calib", package="laGP")
See the package documentation. A pdf version of the reference manual, or help pages, as also available. The help pages can be accessed from within R. Try starting with:
R> help(package=laGP)
R> ?laGP # follow the examples
R> ?aGP # follow the examples - this is the main workhorse

References

Emulating satellite drag from large simulation experiments (2019) with Furong Sun, Benjamin Haaland, Earl Lawrence and Andrew Walker. SIAM/ASA Journal on Uncertainty Quantification, 7(2), pp. 720-759; preprint on arXiv:1712.00182
UChicago RCC did a puff piece on laGP for large-scale computer experiments. RCC resources have played an integral role in laGP development and large-scale application.
Calibrating a large computer experiment simulating radiative shock hydrodynamics (2015) with Derek Bingham, James Paul Holloway, Michael J. Grosskopf, Carolyn C. Kuranz, Erica Rutter, Matt Trantham, R. Paul Drake; Annals of Applied Statistics, 9(3), pp. 1141-1168; preprint on arXiv:1410.3293
laGP: Large-scale spatial modeling via local approximate Gaussian processes in R (2016); Journal of Statistical Software, 72(1), pp. 1-46; provided as a vignette/tutorial above
Local Gaussian process approximation for large computer experiments (2015) with Dan Apley; Journal of Computational and Graphical Statistics, 24(2), pp. 561-578; preprint on arXiv:1303.0383
Modeling an augmented Lagrangian for blackbox constrained optimization (2016) with Genetha Gray, Sebastien Le Digabel, Herbie Lee, Pritam Ranjan, Garth Wells and Stefan Wild; Technometrics (with discussion), 58(1), pp. 1-11; preprint on arXiv:1403.4890
Bayesian optimization under mixed constraints with a slack-variable augmented Lagrangian (2016) with Victor Picheny, Stefan Wild and Sebastien Le Digabel; preprint on arXiv:1605.09466
Massively parallel approximate Gaussian process regression (2014) with Jarad Niemi and Robin Weiss; Journal of Uncertainty Quantification, 2(1), pp. 564-584; preprint on arXiv:1310.5182
Speeding up neighborhood search in local Gaussian process prediction (2016) with Benjamin Haaland; Technometrics, 58(3), pp. 294-303; preprint on arXiv:1409.0074