qmix and idqn agents provided in the scripts directory do not learn

jianzhnie / deep-marl-toolkit

MARLToolkit: The Multi-Agent Rainforcement Learning Toolkit. Include implementation of MAPPO, MADDPG, QMIX, VDN, COMA, IPPO, QTRAN, MAT...

https://jianzhnie.github.io/llmtech/

Apache License 2.0

109 stars 13 forks source link

qmix and idqn agents provided in the scripts directory do not learn #7

Open Timob12 opened 1 year ago

Timob12 commented 1 year ago

Hello, when executing the scripts provided in the scripts directory, my agents do not learn. The training and testing win rate is 0.0 in all cases. I also switched the SMAC map to "3m" and changed difficulty from "7" to "1" to check if that caused the problem, but the agents still do not learn. Is this a problem with the scripts or is there a bug in the implementation? I would like to use this toolkit for my own project, but for that I need to be sure that the implementation works.

jianzhnie commented 1 year ago

https://wandb.ai/jianzhnie/StarCraft-II/runs/2rez7iej?workspace=user-jianzhnie

jianzhnie commented 1 year ago

this is my train log, which seems the test rewards is ok

Timob12 commented 1 year ago

I am sorry but I cannot see your log files, I do not have access to them on w&b. I have not logged things with w&b, I ran it in offline mode. Other than that I just ran your main_qmix.py without any modifications. You might want to add a precise installation guide to your repo because I also encountered some problems with your setup.py which does not contain the "configs" directory by default.

jianzhnie commented 1 year ago