๐Ÿ“ Flatland Community Prize

:hourglass: Deadline Extended to 4th November and Prizes Doubled!

A real-world problem like Flatland inspires niche and novel approaches to find better solutions, and this competition is constantly pushing everyone to get creative and learn from each otherโ€™s performances.

To increase collaboration and novel problem-solving, AIcrowd is excited to announce the Flatland Community Prize:

Create any form of work that contributes towards a better understanding of the Flatland environment by November 4th to win up to 500 CHF!

:memo: CONTRIBUTIONS

Your contribution can be in terms of:

  • Notebooks (runnable online on Colab or Binder)
  • Data visualizations
  • Custom observation builders (see the doc)
  • Articles
  • Videos
  • Anything else

There are no constraints on how you wish to present your ideas! You can take inspiration from some works created by previous participants:

You can also check out these Master Thesis written in the context of the 2019 challenge:

Of course we donโ€™t expect this level of depth - although we donโ€™t forbid it either :stuck_out_tongue_winking_eye:

:trophy: PRIZES

  • 1st place: 500 CHF
  • 2nd place: 300 CHF
  • 3rd place: 200 CHF

:spiral_calendar: NEW DEADLINE

4th November 2020, 23:55 UTC

:balance_scale: TERMS

  1. You must post your explainers on the challenge forum with the tag โ€˜explainerโ€™, introducing and linking to your contribution.
  2. The prizes will be awarded as per the discretion of the organizers, and the popularity of the post in the community (based on number of likes :heart:) - so share your post widely to spread the word!
  3. You can make multiple submissions, but youโ€™re only eligible for the prize once.
  4. Your work needs to be published under a license and on a platform that allows other participants to use it. For example, code should be provided under an open-source license, and articles should be readable freely.

Weโ€™re looking forward to see what you come up with!

4 Likes

:memo: Flatland Community Prize :bulb: Contribution Ideas

Best solutions to a problem like Flatland are found in collaborations and learnings from multiple approaches. To facilitate sharing of ideas and discussions, we invite you to share tools & tips, resources & approaches you follow with the Flatland community. :writing_hand:

:steam_locomotive: Flatland advances state-of-the-art work in multi-agent reinforcement learning, and your contributions can play a key part in that advancement!


If you are confused about what kind of Contributions you can make, we are listing down some ideas in this doc that you can follow. :raised_hands:

We hope they inspire you to share your ideas and insights with the Flatland community!


Simple explainer videos

  • Flatland is a complex problem that may be hard to get into. We have worked hard to revamp its documentation, create new starter kits and baseline examples, but it is still challenging to make your first steps with this environment!
  • We would welcome simple videos that give a high-level overview of this problem, and that walk newcomers through their first experiments with it.

Investigate the use of new methods

  • We have released a number of baselines that show results using methods such as PPO, Ape-X, imitation learning algorithms etc.
  • There are still many interesting methods that could be explored! It would be interesting to investigate the performance to see which one are the most promising:
    • PPG is a brand new improvement of PPO that shows improved sample efficiency.
    • PLR is another bleeding edge method specifically for procedurally-generated environments. The idea is to find and prioritize the environment seeds that provide the most learning potential.
    • MADDPG is an extension of DDPG for MARL environments. Most of the methods we have investigated so far are designed for single-agent environments, and we simply consider the other agents as being part of the environment. But this is not optimal! MADDPG handles this properly.

Improve the rendering of the environment

1 Like