Participate in the next edition of Neural MMO - the largest RL competition at NeurIPS 2022!
NMMO 2.0 is rebuilt for NeurIPS 2023 with faster performance, new focus on task completion, and a new RL baseline.
📃Project Page - Documentation, API reference, and tutorials.
📹WebViewer - A web replay viewer for our challenge.
Your objective is to train agents to complete tasks they have never seen before against opponents they have never seen before on maps they have never seen before
Reinforcement Learning: Customize the RL algorithm, model, and reward structure.
Curriculum Generation: Design the task generator, task sampler, and reward using Python.
No Holds Barred Track: Bring it on! No restrictions; entrants provide their own compute to win via any way possible - except hacking our servers!
- LLM Agents: Use GPT or local LLMs to generate scripted agents. We're still working on evaluation support for this, see Documentation/Discord for details.
October: Competition launches - warm-up rounds vs. baselines
November-December: Main competition. Task completion evaluated against other participants’ Agent policies. Competition tasks get harder over rounds.
TBA Before NeurIPS: Submissions close, final evaluation of top 16 submissions in each track
NeurIPS: Winners notified!
🏆Awards and Prizes
$20K in prizes sponsored by Parametrix.ai. Winners will co-author the summary manuscript following the competition. Per-track and round split to be announced. Winners for the Reinforcement Learning and Curriculum Generatrion tracks are required to open-source full code for their submissions. This is encouraged but not required for the No Holds Barred track.
Reinforcement Learning Track:
You may modify the model architecture, RL algorithm, and reward function.
You may not alter the training tasks or the sampling order of the training tasks.
You may not precompute large amounts of work, for example, through neural architecture search or massive hyperparameter sweeps that tune to multiple significant digits.
Winners will be required to open-source their code in order to be eligible for a cash prize and co-authorship. We will retrain your submission from scratch for 8 hours on an A100 with at least 12 cores. The compute limit is intended to make this track fair for academic labs and independent researchers.
Curriculum Generation Track:
You may modify the generation and sampling of tasks as well as their rewards.
You may not alter the model architecture or RL algorithm.
You may not precompute and upload a specific set of tasks through large scale simulation.
Winners will be required to open-source their code in order to be eligible for a cash prize and co-authorship. We will retrain the baseline from scratch with your curriculum generator for 8 hours on an A100 with at least 12 cores. The compute limit is intended to make this track fair for academic labs and independent researchers.
No Holds Barred Track:
You may modify the model architecture, RL algorithm, reward function, task generation, task sampling, etc. and are not constrained by compute.
Upload your trained model for evaluation.
Winners are strongly encouraged but not required to open-source their code.
Winners will be required to disclose their general approach in order to be eligible for co-authorship.
General Rules for All Tracks:
- Do not attempt to circumvent the submission limit by making multiple accounts or otherwise.
- Do not interfere with our leaderboard by uploading submissions that violate track-specific restrictions.
- Do not attempt to modify the stats recorded to the leaderboard or surreptitiously determine held-out tasks.
- Do not attempt to access or modify other participants’ submissions
- Do not write code that makes alliances with other participants' policies
- Your participation is at our discretion. Harassing organizers, other participants, or other disruptive or rule-breaking behavior will result in a ban and forfeiture of any and all prizes.
Massachusetts Institute of Technology
Kyoung Whan Choe
Hao Xiang Li
Rose S. Shuman
Lucas de Alcântara