Minor Questions/Remarks about gym-microrts-paper-sb3 HOT 3 CLOSED

Hey! Thanks a lot for the feedback.

Yeah, I saw implementation from the contrib package. Though masking itself seemed to me like not the trickiest part of the implementation. And the very first idea with putting masks into observations dictionary to simply mask out logits in a custom mlp_extractor ... just worked out. That's why I didn't look deeply into MaskablePPO implementation. And because PPO itself in algorithms with such a large number of implementation details, the initial goal was to touch as little of the algo itself as possible (ideally, no changes at all).

Regarding the version, the latest version of gym-microrts==0.5.0. requires gym>=0.20.0 which turned out to be incompatible with stable_baselines3==1.3.0. Previously I used MicroRTS 0.4.3 with SB3 1.3.0 but that required me to reimplement some pieces of the environment logic only available in the latest release. I'm not sure if Gym version is critical for MicroRTS, need to do more experiments with different version. It might be easier to wait for SB3 to add support for the latest Gym though 😀

from gym-microrts-paper-sb3.

for gym, this will probably solve your issue: DLR-RM/stable-baselines3#674 (comment)

we don't support it yet as we want to avoid breaking changes and bad surprises for users.

from gym-microrts-paper-sb3.

Yeah, I saw the thread. Seems like a lot of work, thanks for paying so much attention to details! Ultimately, I've decided to wait when it's merged to update 😀

from gym-microrts-paper-sb3.