Learning By Doing - NeurIPS 2021 Competition

Last edited: 2022-11-29

Our paper in the PMLR Proceedings of the NeurIPS 2021 Competitions and Demonstrations Track describes both Track CHEM and ROBO and how they relate to causality, control theory, and reinforcement learning, and discusses the lessons learnt throughout the competition (bib).

The competition is over – thanks to all participants and congratulations to the prize winners!!! Code and winning solutions can be found in this github repo. The repo includes everything you need to make new submissions and evaluate them locally, so feel free to try out the two tracks of the competition. There’s a demo of the repo in the Track ROBO workshop video below.

Videos from the competition workshop with participant talks and discussion can be found below. They include information that was not available to participants during the competition, so if you want a level playing field you should only consult the tutorial and the starter kit.

demo1 demo2

Controlling a Dynamical System using Control Theory, Reinforcement Learning, or Causality

Control theory, reinforcement learning, and causality are all ways of mathematically describing how the world changes when we interact with it. Each field offers a different perspective with its own strengths and weaknesses.

In this NeurIPS competition, we aim to bring together researchers from all three fields to encourage cross-disciplinary discussions. These can happen during the competition and also afterwards when solutions are being presented. The competition is constructed to readily fit into the mathematical frameworks of all three fields and participants of any background are encouraged to participate.

We designed two tracks that consider a dynamical system for which participants need to find controls/policies to optimally interact with a target process: an open loop/bandit track (CHEM) and a closed loop/online RL track (ROBO).

We hope that the challenge further bridges the gap between control theory, reinforcement learning, and causality. Seeing how the same problem can be tackled in different ways may be a first step towards understanding the reasoning in other communities and learning from each other.

NeurIPS Competition Workshop

Our competition presentation is scheduled for Competition Track Day 2, 8th December, 10:05–10:25 GMT. After our competition presentation (and the announcement of the winners), we invite you to join for our Zoom competition workshop (attendance at the workshop or at the presentation requires a NeurIPS registration. Links for Zoom and Gathertown are on this page of the NeurIPS virtual conference site. Zoom link for the competition presentation and live stream of the competition presentation are at the top of the page. Scroll down to find Breakout: Learning By Doing for Zoom and Gathertown links for the competition workshop).

Track CHEM workshop

Track ROBO workshop

Prizes

For each of the two tracks, the following prizes are awarded:

In case of a tie the prizes are split among the winners. See the terms and conditions for eligibility criteria.

Sponsored by the Department of Mathematical Sciences, University of Copenhagen and the Copenhagen Causality Lab.

Getting Started

The competition is run on CodaLab:

Use our tutorial to learn the technical details of how to participate in the competition and get started. For further background information take a look at the white paper.

Please use the Codalab Fora for questions about the competition:

Important Dates

Track CHEM – Trial Phase Start (trial data, starter kits, tutorial)
July 6th, 10:00 UTC
Track CHEM – Validation Phase Start (main competition phase)
July 10th, 16:00 UTC
Track ROBO – Trial Phase Start
July 15th, 17:00 UTC
Track ROBO – Validation Phase Start
July 29th, 16:00 UTC
Registration Deadline (both tracks)
August 20th, 16:00 UTC
Track ROBO – Selection Phase Start
September 12th, 16:00 UTC
Track CHEM – Selection Phase Start
September 20th, 16:00 UTC
Competition Deadline (both tracks)
September 26th, 16:00 UTC
Technical Description Submission Deadline (both tracks)
October 3rd, 16:00 UTC
NeurIPS Conference and Announcement of Winners
December, 13–14, 2021

Organizors

Contact us via email at LearningByDoing AT math DOT ku DOT dk.


Dominik Baumann
RWTH Aachen University

Tabitha Edith Lee
Carnegie Mellon University

Niklas Pfister
University of Copenhagen

Isabelle Guyon
Université Paris-Saclay, ChaLearn

Søren Wengel Mogensen
Lund University

Sebastian Trimpe
RWTH Aachen University

Oliver Kroemer
Carnegie Mellon University

Jonas Peters
University of Copenhagen

Sebastian Weichwald
University of Copenhagen

Citation

@inproceedings{weichwald2022learning,
      title=, 
      author={Sebastian Weichwald and Søren Wengel Mogensen and Tabitha Edith Lee and Dominik Baumann and Oliver Kroemer and Isabelle Guyon and Sebastian Trimpe and Jonas Peters and Niklas Pfister},
      booktitle={Proceedings of the NeurIPS 2021 Competitions and Demonstrations Track},
      pages={246--258},
      year={2022},
      volume={176},
      series={Proceedings of Machine Learning Research},
      month={06--14 Dec},
      publisher={PMLR},
      url={https://proceedings.mlr.press/v176/weichwald22a.html},
}