Ecole Paper

This repository contains the conda for the Ecole paper, including comparison with Exact combinatorial optimization with graph convolutional neural networks Gasse, Chételat, Ferroni, Charlin, and Lodi (2019) in Advances in Neural Information Processing Systems (pp. 15580-15592).

Setup

To install Ecole and all the dependencies, run

git submodule update --init --recursive
conda env create --name ecole-paper --file environment.yaml
conda activate ecole-paper
conda env update --file "vendor/ecole/dev/conda.yml"
cmake -B ecole_build -S vendor/ecole -D ECOLE_BUILD_BENCHMARKS=ON -D CMAKE_BUILD_TYPE=Release
cmake --build ecole_build --parallel
pip install ecole_build/python
pip install .

Generating the data

For benchmarking the Ecole overhead:

./build/libecole/benchmarks/benchmark-libecole --ipg 375 --nl 100 --seed 42 >> data/benchmark-branching.csv

For benchmarking the observation functions:

python -u -m ecole_vs_gasse.bench_observation --nl 100 --ipg 35 --seed 740 >> data/benchmark-observation.csv

Visualizing the results

The notebook Analysis.ipynb provide code to analyse the results and reproduce the table of the paper.