This is a pytorch implementation of the multi-agent reinforcement learning algorithm, DyMA-CL

Beacuse previous StarCraft II settings enable an agent to attack one of its enemies by choosing one of id numbers, the action spaces of different maps are not the same, so in the original paper, the author design the attack action is to choose one of the grid units by dividing the battlefield into several grids.

We don't want to waste time modifying environment code, so we just run the algorithm on map '8m' and '3m', and for '3m', we pad its action space with 5 zeros to make sure it has the same action space dimension as '8m', but we forbid agents to choose these 5 actions.

Because we use zero-padding, there is a problem for transferring model from '3m' to '8m'. In '3m', the last 5 actions have never be chosen, so their Q-value is far from true value, when we use the model of '3m' on '8m', we find the performance of transferring is negative. Therefore, we transfer the model from '8m' to '3m' to avoid this problem. We first train the model on '8m' for a while, and then transfer the model to '3m', and we compare the result between transferring and not transferring.

$ python main.py

Directly run the main.py, then the algorithm will start training on map '3m', using the pretrained model on map '8m'. If you want to train from scratch, you should use it to train on '8m', then move the model of '8m' to ./model/dyma/3m/ and rename the model.

We independently train DyMa for 8 times and take the mean of the 8 independent results.