hmer

Overview

The goal of hmer is to make the process of history matching and emulation accessible and easily usable by modellers, particularly in epidemiology. The central object of the process is an Emulator: a statistical approximation for the output of a complex (and often expensive) model that, given a relatively small number of model evaluations, can give predictions of the model output at unseen points with the appropriate uncertainty built-in. Using these we may follow a process of ‘history matching’, where unfeasible parts of the parameter space are ruled out. Sampling parameter sets from the remaining region allows us to train more accurate emulators, which allow us to remove more of the space, and so on. The hmer package contains tools for the automated construction of emulators, visualisations for diagnostic checks and exploration of parameter space, and a means by which new points can be proposed.

Installation

You can install the development version of hmer from GitHub with:

# install.packages("devtools")
devtools::install_github("andy-iskauskas/hmer")

Example

The three core functions of the package are called below, using built-in toy data.

library(hmer)
#> Registered S3 method overwritten by 'GGally':
#>   method from   
#>   +.gg   ggplot2
## Train a set of emulators to data
ems <- emulator_from_data(input_data = SIRSample$training,
                          output_names = names(SIREmulators$targets),
                          ranges = list(aSI = c(0.1, 0.8), aIR = c(0, 0.5), aSR = c(0, 0.05)))
## Perform diagnostics on the emulators
validation <- validation_diagnostics(ems, SIREmulators$targets, SIRSample$validation, plt = FALSE)
## Propose new points from the emulators
new_points <- generate_new_design(ems, 50, SIREmulators$targets)

Learning Emulation and History Matching

There is a wealth of published information on Bayes Linear emulation, history matching, and the more general framework of uncertainty quantification, upon which this package is based. The easiest way to learn how to use the hmer package, however, is to look through the vignettes within.

browseVignettes("hmer")
vignette("low-dimensional-examples", package = 'hmer')

• Low-dimensional examples low-dimensional-examples introduces the basics of emulation and history matching and how to use hmer in some low-dimensional toy models;

• Demonstration demonstrating-the-hmer-package serves as a broad overview of most of the functions in the package;

• Stochastic and Bimodal Emulation stochasticandbimodalemulation introduces the basics of dealing with stochastic systems, and identifying bimodality;

• The “Emulation Handbook” emulationhandbook details some common problems and considerations that occur when using the framework, and serves as a broad FAQ for problems encountered.