Mlegp Command Line Interface

ccs, [email protected] 09.05.2013

A Rscript driven command line interface to the wonderful Mlegp package. http://cran.r-project.org/web/packages/mlegp/index.html

Contents

MlegpInter

A small R library (mlegpInter) with some helper functions to make predictions from an mlegp GP list (in ./R-pkg)

Scripts

• mlegp-interactive-train.R: train a set of GP's on a multivariate data set with multivariate output.
• mlegp-interactive-predict.R: make predictions (mean, variance, covariance) from the trained GP's produced by "mlegp-interactive-train", reads points in parameter space from stdin, prints output to stdout
• mlegp-interactive-implaus.R: output the joint and independent implausibility for a given set of trained GP's. Needs a file containing the observed means and std.errors.

Depends

• R (tested on 2.15.2, 2.15.3), Rscript
• R packages: mlegpFull (get this from the maintainer), adapt, optparse

Installing

Install the R Package

• The library depends upon the following R packages: optparse, mlegp, adapt, mlegpFull the latter can be obtained directly from the mlegp maintainer. You can install any missing dependencies via R by using the install.packages command.

cd ./R-pkg

• tell R to build the library and install it R CMD INSTALL .

• you can access the lib in R with: library(mlegpInter) but there's basically no need to.

Install the Scripts

• copy the contents of scripts to somewhere in your path, or leave them where they are.

Usage

• see /example/mw1-5param-example/README.md for a full example

Plotting Main Effects

If you've trained a multivariate GP using the melgp-interactive-train script you can now make plots of the main effects in the true and pca basis (of the outputs). There's no direct command line support for this but it's easy to do from an R terminal. The functions provided by this library just wrap the mlegpFull

• plot.main.effects.pca plots the effects in the pca basis
• plot.main.effects.true plots the effects in the true basis

both functions also return a list of matrices of the main effects for each observable that can be written to file etc.

Suppose that you've saved the data from the training process into "trained-emu-save.dat", to plot the main effects in the PCA basis do the following in an R process:

library(mlegpFull)
library(mlegpInter)
load("trained-emu-save.dat")
main.pca.table <- plot.main.effects.pca(fit.pca)

This will produce a set of graph panels on the default device and main.pca.table will be init to a list of the main effects. Each list item is a matrix of the effects for a given PCA observable with the first column as the arbitrary index on the parameter space and then each column giving the predicted main effect for that parameter.

The process is the same for plotting the main effects in the true basis but one needs to supply the training scale info. This saved by default as training.scale.info in the file created by the mlegp fitting process.

library(mlegpFull)
library(mlegpInter)
load("trained-emu-save.dat")
main.pca.table.true <- plot.main.effects.true(fit.pca, train.scale.info=training.scale.info)

The resulting table can be saved to disk etc.

TODO

• add a command line interface to the SA code
• check the design/model-data for actually being numeric and complain appropriately