The issue is explained in PR #111. After careful consideration and discussions with @samvelyan, I have decided not to make any changes to the code. This is because the issue will have minimal impact on the training performance of most multi-agent reinforcement learning (MARL) algorithms, and many models have already been trained and saved using the current observation space.
Instead, I have added some comments in the starcraft2.py to help users align the unit type id in enemy and ally features if they are going to decompose the observation or state space manually to customize their models.
samvelyan
commented
1 year ago
The issue is explained in PR #111. After careful consideration and discussions with @samvelyan, I have decided not to make any changes to the code. This is because the issue will have minimal impact on the training performance of most multi-agent reinforcement learning (MARL) algorithms, and many models have already been trained and saved using the current observation space.
Instead, I have added some comments in the
starcraft2.pyto help users align theunit type idin enemy and ally features if they are going to decompose the observation or state space manually to customize their models.