Repo of the CP22 paper:

Mohit Kumar, Samuel Kolb, Tias Guns: Learning Constraint Programming Models from Data Using Generate-And-Aggregate. CP 2022: 29:1-29:16

PDF available at: https://drops.dagstuhl.de/opus/volltexte/2022/16658/pdf/LIPIcs-CP-2022-29.pdf

For questions please contact Prof. Tias Guns

Paper history

We first created prototype code for the PTHG 21 Constraint Acquisition Challenge: https://freuder.wordpress.com/progress-towards-the-holy-grail-workshops/pthg-21-the-fifth-workshop-on-progress-towards-the-holy-grail/

We then refactored and extended the parts that we felt would make for an interesting paper. That is provided in this repo.

Requirements

The core part of CountCP (learn.py) uses the constraint solving library CPMpy. At submission time CPMpy was at v0.9.7, but it could be that we were using an older version.

I hope everything continues to work with the latest CPMpy version, if not let us know so we can fix it.

We also use the SymPy library for symbolic expressions.

Repo structure

• instances/ contains the 17 original instances from the PTHG21 challenge including our post-hoc interpretation of the ground truth model expressed in CPMpy (based on the names of the problems shown in the PTHG21 slides, copied into instances/problems.txt. We do not know the real ground-truth models), plus our own simple nurse rostering instance generator that outputs PTHG21 formatted data.
• README.md this file, send us updates if you feel something is missing
• cp2022_experiments.py tools for running and measuring the experiments in the CP22 paper. Check out the main part to see what arguments are sensible (e.g. graph, sudoku, queens, magic and nurses)
• instance.py class(es) to help manage the PTHG21 instances
• learn.py This is the core of the code and contains the learnin logic
• musx.py CPMpy MUSX implementation from its advanced example directory
• results.py helper code to create visualisations from the experiment logs