GADGET: Gaussian Process Approximations for Designed Experiments

The GADGET package computes near-optimal Bayesian experimental designs using Gaussian process optimization. At its core is the ability to calculate static designs that maximize a design criterion that may be either deterministic or stochastic. In particular, stochastic design criteria could be a Monte Carlo estimator of an expected utility based on MCMC posterior draws. GADGET utilizes the algorithm proposed by Weaver et al. (2016) and performs Gaussian process validation using the statistics introduced by Bastos and O’Hagan (2009). The parallel package is integrated to parallelize the evaluation of the user’s design criterion. Additionally, GADGET has wrapped the optimization into a sequential routine to perform sequential computer experiments that automatically call simulator code that is available in R.

Installation

To install GADGET from github, use the install_github function from the devtools package.

install.packages('devtools')
devtools::install_github('isaacmichaud/GADGET')

References

Bastos, L. S., & O’Hagan, A. (2009). Diagnostics for gaussian process emulators. Technometrics, 51(4), 425–438.

Weaver, B. P., Williams, B. J., Anderson-Cook, C. M., Higdon, D. M. (2016). Computational enhancements to Bayesian design of experiments using Gaussian processes. Bayesian Analysis, 11(1), 191–213.