koulanurag / minimal-marl Goto Github PK

Hi, I have a question in your qmix implementation. I do not understand why you sum the losses at every chunk (line 151) whereas in the original paper we sum over the batch only (equation 6). Why don't we just compute the loss of the final value of the sequence instead ?
Thanks

Hello, thanks for the work. From your results on wandb, it seems like QMIX is a bit unstable/suboptimal. Any idea why?

Additionally, do you have some training results for IDQN? Thanks again

Hi, in this line

Hi when I run python idqn.py, or vdn.py or qmix.py, there is a rumtime error.
"RuntimeError: cannot perform reduction function argmax on a tensor with no elements because the operation does not have an identity"
It is caused by "action[~mask] = out[~mask].argmax(dim=2).float()", I suppose it is because out[~mask] is empty when mask=True.
Please have a look, thanks.