benoitsteiner / ml4co-competition Goto Github PK

Official website: competition guidelines, team registration, rules, and leaderboard.

This repository contains the base code that supports the competition, as well as some code examples and baseline implementations for each of the three decision tasks that participants can compete in (primal, dual, config). It is organized as follows

instances/    -> the datasets
common/       -> the common code base, i.e., environments, rewards and evaluation scripts
submissions/  -> the team submissions
  example/    -> an example submission
singularity/  -> the singularity image and scripts of our evaluation pipeline

• Getting Started: get the data, implement and evaluate your agent, make a submission.

• Data description: the three datasets (item_placement, load_balancing, anonymous) and the data files.

• Tasks description: how we implemented each task (primal, dual, config).

• Evaluation pipeline: make sure your code installs and runs within our pipeline before you submit.

• Evaluation platform: the hardware and software specifications of the platform your code will be evaluated on.

• APIs: Ecole - SCIP - PySCIPOpt

During the competition we will evaluate submissions on a weekly basis, and update our online leaderboards while the competition is running. Participants do not have to send a submission every week, but are encouraged to submit regularly to make the competition a live event. In order to prevent test overfitting, those intermediate evaluations will be performed on a fixed subset of the test set (20%), while only the final evalaution that will tell the winners will be performed on the entire test set.

We provide an official support to participants via the Github discussions feature. Please direct any technical or general question regarding the competition there, and feel free to answer the questions of other participants as well. We will not provide a privileged support to any of the participants, except in situations where it concerns information about their submission which they do not want to share. To contact us directly, use the competition's email adress: [email protected] .

We will not run the training of your ML models. Please send us only your final, pre-trained model, ready for evaluation.

We only offer support for Linux and MacOS, the Windows platform being currently unsupported by Ecole.

We thank Compute-Canada, Calcul Québec and Westgrid for providing the infrastructure and compute ressources that allow us to run the competition.