MIT + AIcrowd Reinforcement Learning Challenge on Neural MMO
Joseph Suarez
As a fellow reinforcement learning practitioner, I invite you to participate in the MIT Neural MMO challenge on AIcrowd.
Neural MMO is a massively multiagent platform for embodied intelligence research (appearing in NeurIPS 2021). This new frontier for reinforcement learning requires policies capable of robust reasoning within large populations and multitask optimization over extended time horizons.
Submit AI agents to compete against each other in procedurally generated MMORPG-inspired virtual worlds. The challenge does not require large-scale hardware -- we provide a google colab starter notebook that will walk you through your first submission in only a few minutes.
The challenge runs until the end of November and features 3 rounds. We look forward to seeing your innovative approaches succeed in this new setting. If you have any questions or queries regarding the challenge, please reply directly to this email or join us on Discord.
- Joseph Suarez, creator of Neural MMO
You can find detailed information about the challenge below:
Neural MMO - Mimicking the real-world through massively multiplayer games
Neural MMO features hundreds of concurrent agents, multi-thousand-step time horizons, high-level task objectives, and large, procedurally generated maps.
MMORPGs (Massive Multiplayer Online Role Playing Games) such as World of Warcraft or Runescape are some of the best game proxies for the real world. Unlike the arcade and round-based games typically used in reinforcement learning research, MMOs are complete macrocosms featuring thousands of agents, diverse skilling systems, global economies, and ad-hoc high stakes single and team-based conflict. They simulate persistent worlds that support rich player interactions and a wide variety of progression strategies.
These properties seem important to intelligence in the real world, and the objective of the Neural MMO Challenge is to spur agent-based research on increasingly general environments -- and to do so without prohibitive computational costs!
Making the environment accessible to all
Full-scale MMOs are among the most complex games developed, are not typically open-source, and are not computationally accessible to most researchers.
We have therefore chosen to build Neural MMO from the ground up to capture key elements of the genre while remaining efficient for training and evaluation. You can train effective policies with a single GPU.
Neural MMO is fully open-source and includes scripted and learned baselines with all associated code. Check out the Google colab notebook here - you can make your first submission to the challenge within a few minutes! Scripted submissions and imitation learning are also fair game.
Evaluation against other participants
MIT has created this challenge in partnership with AIcrowd to promote embodied intelligence research on increasingly complex and general environments.
Participants submit agent policies to be evaluated in a shared virtual world against other submissions. You earn points by completing high-level tasks in the presence of dozens to hundreds of opponents not seen during training.
Agents are ranked according to their performance against all other submissions. Individual scores for each submission are calculated as below:
(more information about the evaluation here)
The challenge runs until the end of November 2021. Round 1 ends in 61 days. This is our first publicized announcement -- starting now will put you on fair ground with other participants.
The top three teams from each round will be invited to contribute material detailing their approaches and will be included as authors in the summary manuscript at the end of the competition.
See you on the leaderboard!