search

google-deepmind / meltingpot

A suite of test scenarios for multi-agent reinforcement learning.
Apache License 2.0
582 stars 118 forks source link

Update pettingzoo stable-baselines3 training script & compatibility wrapper #117

Closed elliottower closed 1 year ago

elliottower commented 1 year ago

I've updated the pettingzoo stable-baselines3 (SB3) training script to work with the current pettingzoo API and use gymnasium rather than gym. I had to restructureexamples/pettingzoo/utils.py but it now aligns better with other pettingzoo compatibility wrappers. I'm currently working on creating a PR in https://github.com/Farama-Foundation/Shimmy so that the conversion code can be more easily found and better documented. This will remove the need for utils.py, but I am keeping it as a placeholder for now.

Note that this relies on a feature branch of SB3 which is not yet merged: https://github.com/DLR-RM/stable-baselines3/pull/1327 This also relies on a PR of supersuit which is not yet merged: https://github.com/Farama-Foundation/SuperSuit/pull/206

Once those two are merged I can update requirements.txt to reflect it, and/or modify install_extras.sh to include stable-baselines3 and supersuit.

elliottower commented 1 year ago

Note: this PR will be updated to use shimmy rather than utils.py once https://github.com/Farama-Foundation/Shimmy/pull/39 is merged. I am thinking it's probably best to also do a separate tutorial file (or other documentation) to demonstrate using melting pot with pettingzoo (using shimmy), and either mark that the SB3 tutorial is not currently working, or wait on merging this PR until SB3 is updated to support the updated gymnasium API (see https://github.com/DLR-RM/stable-baselines3/pull/780)