doit: Bayesian Computation using Design of Experiments-based Interpolation Technique in R

Overview

This R package implements the design of experiments-based interpolation technique (DoIt, Joseph 2012) for approximate Bayesian computations.

The method uses evaluations of an unnormalised density at a space-filling design of parameter values. Normalisation is achieved by approximating the target density by a weighted sum of Gaussian kernels centered on the design points.

DoIt approximates the joint density, marginal densities, as well as expecations and variances. The package contains functions to optimally choose additional design points, and to calculate the optimal kernel width by cross validation.

Figure: DoIt approximation of a complicated 2-dimensional density. See vignette('doit_2d') for details.

Installation

remotes::install_github('sieste/doit', build_opts=NULL)

To install the package without using remotes, run the following shell commands:

git clone git@github.com:sieste/doit.git
cd doit
R CMD build .
R CMD INSTALL doit_*.tar.gz

Vignettes

The usage of the package is documented in 2 vignettes, where results from the original papers are reproduced.

vignette('doit_1d') # 1d example from Joseph (2012)
vignette('doit_2d') # 2d example from Joseph (2012)

References

Joseph (2012) Bayesian Computation Using Design of Experiments-Based Interpolation Technique, Technometrics, 10.1080/00401706.2012.680399

Joseph (2012) A Note on Nonnegative DoIt Approximation, Technometrics, 10.1080/00401706.2012.759154