fd.maxent: Solve Maximum Entropy Problems in R

This R-Package implements several algorithms to solve constrained maximum entropy problems (currently discrete problems only) and provides several applications and use cases to explore the pros and cons of the algorithms.

Installation

You need R or RStudio to use the package.

To install fd.maxent, you need the packages remotes, rmarkdown and knitr. These get installed with

install.packages("remotes")
install.packages("rmarkdown")
install.packages("knitr")

Then you can execute

remotes::install_github('kweinert/fd.maxent', build_vignettes=TRUE)

Typically, you will be asked during the installation to install or update additional packages. It is recommended to choose the option "CRAN packages only" in this case.

You can test whether the installation was successfull by running the unit tests. This requires the package tinytest:

install.packages("tinytest")
library(fd.maxent)
mep_test()

Usage

The main functions in the package start with mep_. Two simple problems aim to explain the idea and can be viewed with:

vignette("simple", package="fd.maxent")

When the problems become more complex, some help from mapping the problem into a numerical formulation is desireable. The functions set, variab, index and vec2df are the approach taken in this package and introduced in the mapping vignette.

Examples

This repository digests some papers and books I read on maximum entropy. You can find the reading list including reviews on wyrms.de and in the literature.bib file in the vignettes folder.

Here is a list of worked examples:

• vignette("europawahl19") is an example which combines several surveys to estimate voting behaviour with respect to socialdemographic variables.
• vignette("balancing") is an example which uses the maximum entropy principle to mitigate the survey reponse bias.

Support

Please reports bugs and ideas via the Issue-Board. You can contact me also directly via Mastodon

License

This project is free software as in GPL-3. See LICENSE for details.

Roadmap

Next steps are

• synthetic balancing example which reweighs a survey due to known population parameters to remedy nonresponse-bias
• inequality constraints
• logistic regression example (Titanic dataset)
• Poisson regression example
• lotto example that estimates pick probabilities from observed quotas
• document the LP approach via a vignette
• Implement interior point algorithm
• Implement solver using the highs package
• Comparison of algorithms

Please note that this is a one-man side-project, so don't expect anything soon.