Extensible Combinatorial Optimization Learning Environments


Keywords
combinatorial-optimization, gym, markov-decision-processes, ml, scip
License
BSD-3-Clause
Install
pip install ecole==0.8.0.dev0

Documentation

Ecole logo

Ecole

Test and deploy on Github Actions

Ecole (pronounced [ekɔl]) stands for Extensible Combinatorial Optimization Learning Environments and aims to expose a number of control problems arising in combinatorial optimization solvers as Markov Decision Processes (i.e., Reinforcement Learning environments). Rather than trying to predict solutions to combinatorial optimization problems directly, the philosophy behind Ecole is to work in cooperation with a state-of-the-art Mixed Integer Linear Programming solver that acts as a controllable algorithm.

The underlying solver used is SCIP, and the user facing API is meant to mimic the OpenAi Gym API (as much as possible).

import ecole

env = ecole.environment.Branching(
    reward_function=-1.5 * ecole.reward.LpIterations() ** 2,
    observation_function=ecole.observation.NodeBipartite(),
)
instances = ecole.instance.SetCoverGenerator()

for _ in range(10):
    obs, action_set, reward_offset, done, info = env.reset(next(instances))
    while not done:
        obs, action_set, reward, done, info = env.step(action_set[0])

Documentation

Consult the user Documentation for tutorials, examples, and library reference.

Discussions and help

Head to Github Discussions for interaction with the community: give and recieve help, discuss intresting envirnoment, rewards function, and instances generators.

Installation

Conda

Conda-Forge version Conda-Forge platforms
conda install -c conda-forge ecole

All dependencies are resolved by conda, no compiler is required.

Pip wheel (binary)

Currently unavailable.

Pip source

PyPI version
Building from source requires:
pip install ecole

Other Options

Checkout the installation instructions in the documentation for more installation options.

Related Projects

  • OR-Gym is a gym-like library providing gym-like environments to produce feasible solutions directly, without the need for an MILP solver;
  • MIPLearn for learning to configure solvers.

Use It, Cite It

Ecole publication on Arxiv

If you use Ecole in a scientific publication, please cite the Ecole publication

@inproceedings{
    prouvost2020ecole,
    title={Ecole: A Gym-like Library for Machine Learning in Combinatorial Optimization Solvers},
    author={Antoine Prouvost and Justin Dumouchelle and Lara Scavuzzo and Maxime Gasse and Didier Ch{\'e}telat and Andrea Lodi},
    booktitle={Learning Meets Combinatorial Algorithms at NeurIPS2020},
    year={2020},
    url={https://openreview.net/forum?id=IVc9hqgibyB}
}