In libKriging/dolka: Utility functions for Bayesian optimization

R package dolka

dolka is a transitional R package aimed at helping to use rlibkriging as an alternative to DiceKriging when doing Bayesian optimization with DiceOptim or when performing kriging inversion with KrigInv.

The name is built as an acronym for DiceOptim with the LibKriging Alternative.

The idea is to rely on R methods, be they S3 or S4, to allow an easy switch from an object with class "km" to an other object with a different class for instance "KM"in the rlibriging R package. The key methods to be enhanced are the predict and update methods since most of the tasks in Bayes optimization and kriging inversion rely on these methods. The method simulate is also important. In a first step, these methods will be re-implemented in dolka based on existing source codes from the packages cited, in view of overloading the existing methods for the "km" class. Then, we will suggest some light refactoring to the authors of DiceOptim, KrigInv and maybe to those of other packages using DiceKriging, if they want their package to allow the use of both DiceKriging and libKriging via its R interface rlibkriging.

Of major importance in Bayesian optimisation, the predict method should be able to provide the derivative of the kriging mean and the kriging covariance w.r.t. one or several "new" design points. This is made possible by using the deriv argument in the predictmethod for the class "km"provided by dolka, which overloads the existing method when dolka is attached.

News

• The genoud_cache and optim_cache functions are wrapers for the optimization function genoud from the rgenoud package and optim from stats. In both cases the function to be optimized and its gradient are provided as one single function returning a named list with the elements objective and gradient as is the rule in the nloptr package. Compared to using two separated functions, this usually allows some savings due to the use of auxiliary variables. To stick to the use of genoud and optim the two required functions fnand gr are created "on the fly" in the wrapers, the gradient being "cached" in an environment.

• The Bayesian optimization criteria to be optimized at each step (Expected Improvement, ...) are and will be provided in versions _with_grad returning the list with objective and the gradient as described above. The computation of the gradient is made optional thanks to the deriv logical argument, and the functions can return simply the numeric value of the objective, possibly with a gradient attribute. This versatility allows the use of different optimization functions e.g. taken from nloptr. For now the Expected Improvement EI and the multi-points a.k.a q-point Expected Improvement qEI.

INSTALLATION

For now, we do not provide precompiled binary versions (.zip for Windows). These may be provided soon in the Releases tab.

Simple installation using remotes or devtools

Since dolka does not for now need compilation (it doe not embed C or Cpp program files), the remotes package can be used.

remotes::install_github("libKriging/dolka")

If you have the Rtools installed, you can also rely on devtools package

devtools::install_github("libKriging/dolka")

In both cases many options are available e.g. to select a specific branch or commit. See the manuals of remotes or devtools for more information.

Clone, build and install

Cloning the repository

If you do not have yet a local dolka repository, use git clone to clone the dolka repository

git clone https://github.com/liKriging/dolka

This will create a dolka sub-directory of the current directory, i.e. the directory from which the git command was issued.

Installation on Unix and MacOs systems

With these systems you can install a package from its source. Move to the parent directory of your cloned repository and use the following command from a terminal to create a tarball source file

R CMD build dolka

This will produce a so-called source tarball file dolka_x.y.z where x, y and z stand for the major, minor and patch version numbers. Then you can install from a command line

R CMD INSTALL dolka_x.y.z.tar.gz

Of course if you are using the RStudio IDE you can install the source tarball using the menu Tools/Install Packages... Note that a source tarball file is called a Package Archive File in RStudio.

Installation on Windows systems

Provided that you have the Rtools installed, you can use the same commands as Linux and MacOS sytems.

Moreover you can create a so-called precompiled binary file using

R CMD INSTALL --build dolka_x.y.z.tar.gz

This will create a dolka.zip file that can be used on a Windows plateform which may not be equipped with the Rtools. For instance, with RStudio you can use the menu Tools/Install Packages and select Install from: to install a precompiled .zip file.

libKriging/dolka documentation built on April 14, 2022, 7:17 a.m.