MARLO 2018

Multi-agent Reinforcement Learning in Minecraft

The competition

MARLO 2018 is an AI contest where teams from all over the world are asked to develop algorithms for single-agent and multi-agent reinforcement learning games.

Join us and participate in CrowdAI’s Project Marlo, where we will be developing reinforcement learning algorithms to guide multiple Minecraft agents through a series of minigames, which include running mazes, herding livestock, and crossing obstacles. If we are one of the top teams, we could get cash prizes, opportunities to present our project at research conferences, and more! The competition has already started, so apply now!

Here is an example of last year’s Malmo project in action.

If you are interested in joining us, please do the following:

- Understand lectures 10-11 of Berkeley’s CS188 course or equivalent (do additional research to supplement understanding if needed):
- Develop a reinforcement learning algorithm to tackle openAI Gym’s cartpole-v1, and be prepared to explain your code during a brief interview.
- Print or display:
  1) Avg. reward over the last 100 consecutive episodes, and
  2) How many episodes before solve (how many episodes before you encounter 100 consecutive episodes with an average score of at least 195. Refer to the environment's documentation for more details)
  3) Any other information you deem important, such as a cumulative reward graph
- Email with [Marlo] in the subject. Link your CartPole git repository and optionally include a resume and/or cover letter.

Useful information:
- Admission deadline: Oct 13
- Project timeline: now - Oct 21
- Link to competition page:
- Software stack (subject to change): Python, Chainerrl, NumPy, Gym, Git.

If you have any questions or concerns, feel free to contact Justin Wang, our competitions manager, at

Happy coding!

Join our team

Give us your information and tell us something about you!

Davis, CA 95616

Your message has been sent. Thank you!