Support for PettingZoo? about epymarl HOT 3 CLOSED

I do not believe that supporting easy interdependency with PettingZoo, one of the most well-supported libraries for MARL environments is "redundant". Requiring potential users to dabble in low-level environment manipulation to conform to a Gym specification and not having comprehensive documentation is problematic for wider adoption, in my view.

from epymarl.

AFAIK, cooperative MARL has the same interface as OpenAI's Gym (which is the bigger project here). You can easily define a tuple of observations (one tuple element for each agent), actions, etc - and a single reward (because it's cooperative). That's identical to Gym (which we currently support). I don't see the reason for supporting a redundant API like PZ's but we will be happy to merge a wrapper if someone wants to contribute it.

from epymarl.

Hi! Sorry for the late follow-up.

There have been several difficulties of an improved integration of PettingZoo due to backwards incompatible changes introduced in later gym and more recently gymnasium versions which PettingZoo relies and that were not compatible with EPyMARL and several of our environments that relies on the outdated gym==0.21 version.

That being said, I am very glad that we just published a big update to EPyMARL including lots of new features and a move to gymnasium from gym (as well as support for training in non-common-reward games, weights and biases logging, and more)! This required not just changes to EPyMARL but also all our environments (esp. RWARE, LBF, matrix games, SMAClite) to continue support of these envs with EPyMARL. As part of this move to gymnasium, we could now easily integrate PettingZoo environments. I have tested it with MPE environments of PettingZoo as well as some simpler environments and it seems to work well. Note, that continuous control tasks will not be supported.

I will close this issue since PettingZoo integration now works, but please let us know if there are any outstanding issues!

from epymarl.