R Packages' Life
I wrote a number of R Packages with a majority of "private" packages, and some have been used for more than 15 years. Unfortunately, I can not maintain packages or participate to their development unless I have some kind of financial support for that.
In the last decade I have had the opportunity to participate to the developpement of some packages that are now released on CRAN, working with different co-authors and contributors. This is always an interesting and challenging adventure. Here are a few words about Renext, DiceKriging and kergp .
CRAN Package Renext
Renext is the result of a collaboration with IRSN Behrig. IRSN provided the financial support and the general conception, due to Claire-Marie Duluc, Lise Bardet, Laurent Guimier et Vincent Rebour. We took over writing R codes and the documentation. Renext is available on CRAN since 2010. Some evolutions and extra documentation are available on this IRSN site.
Renext is devoted to the Peaks Over Threshold method (POT) as called "méthode du renouvellement" by french-speaking hydrologists, though a more appropriate name would be Marked Poisson Process modelling. The package implements special likelihood maximisation by concentration for several distributions (Lomax, GPD, Weibull, ...) and several frameworks. The package allows using observations resulting from a temporal aggregation as often used in historical information. These include as special cases: block maxima and \(r\)-largest observations. Inside Renext we find: plotting positions for heterogeneous data, a specific test of exponentiality against GPD alternatives, ...
CRAN Package DiceKriging
DiceKriging is a package co-authored with Olivier Roustant, David Ginsbourger. As the name might suggest, the development began within the Dice consortium. The package is released on CRAN since 2010. It includes contributions by Clément Chevalier and by Yann Richet (IRSN) who also implemented Shiny illustrations and provided support for Continuous Integration tools on the GitHub repository. DiceKriging is devoted to kriging tasks as needed in Computer Experiments, including the ubiquitous scatter data interpolation with confidence bands.
DiceKriging has become quite popular over the years especially in the communities of Computer Experiments and Bayesian Optimisation. It has more than 25 CRAN reverse dependent packages.
CRAN Package kergp
kergp is a package co-authored with Olivier Roustant, David Ginsbourger, with contributions by Nicolas Durrande. It is devoted to Gaussian Processes, with emphasis on the choice of covariance kernels. Some kernels such as Matérn-Whittle are provided, but user-defined kernels can as well be coded, either in R or by using compiled code for a greater efficiency. All these known kernels allow to further define composite kernels: radial, tensor product, ... Moreover, kernels are provided for categorical inputs (nominal or ordinal), and we can define composite kernels with mixed inputs: some being continous, some being categorical. Useful applications include kriging with derivatives, and more.
kergp has been proved to be useful in research on kernels and has been used by several young researchers who helped at the developpement, bug fixes and documentation: Michaël Binois, David Gaudrie, Julien Pelamatti. kergp has been created during ReDice consortium. In the years 2016-2020, my own work on kergp has been supported by the chair Oquaido and I co-organised several training sessions for Oquaido members.
