lich14 / cds Goto Github PK

Dear, author:
CDS is good paper and has a good performance on several super hard maps of SMAC. I am confused about the intrinsic reward is implemented in the code.
In predict_net.py line 154, log_p_o (represent log p(o_{t+1}|\tau_{t},a_{t})) is calculated by -1 * F.mse_loss(predict_variable, other_variable, reduction='none'). I want to know why use -1 * F.mse_loss(predict_variable, other_variable, reduction='none') instead torch.log(predict_variable) for calculating log_p_o?

Hello!
Note that the provided config files do not contain 3s_vs_5z and 5m_vs_6m, whose results are included in the paper. Could you provide the config of these 2 maps?

I ran the two environments in the paper that were significantly better than Qplex, and the results were consistent with the paper. CDS is a great job!

Greetings,

Thanks a lot for providing the code.
I was wondering if it is possible for you to share code for running experiments on pacmen_env as well?
Something like main.py that you provided for GRF, but for pacmen_env?

Thank you

Hi authors! I really appreciate that you released the code of CDS. Here I have a question when reproducing the results of CDS in the SMAC environment. I wonder if "qplex_qatten_sc2" is the CDS algorithm for SMAC? And what is "qplex_sc2_ablation"? Thanks!

Greetings！The paper and algo is really helpful to QMIX and QPLEX.
After reading the repo's README, i find it is also said to be possible to implement CDS on MAPPO, but I
have no idea about it.
Could you provide some advice or guide about policy-based algorithm such as MAPPO？
Thanks！

Dear authors,

Thank you for your awesome work.

When I was running experiments for GRF experiment, I noticed a performance discrepancy between the paper(or the figure in the data folder) and the codebase. For example, I was running CDS on Qmix with main.py --config=CDS_QMIX --env-config=academy_3_vs_1_with_keeper and I only got 27.5% (averaged over 5 seeds) test average score after 4M environment timesteps. After I remove the CDS part, the performance drops to 7.5%. So this proves that CDS is indeed very useful in improving the QMIX performance.

However, it would be really helpful if you could advise on the discrepancy in settings (for examples, batch size, number of parallel runs) between this codebase and the paper. Thank you very much.

https://github.com/lich14/CDS/blob/main/CDS_SMAC/QPLEX-master-SC2/pymarl-master/src/run.py#L106
There is no get_unit_dim() function in the standard smac env or qplex-modified smac env.
AttributeError: 'StarCraft2Env' object has no attribute 'get_unit_dim'
just change args.unit_dim = runner.env.get_unit_dim() to args.unit_dim = env_info["unit_dim"]

Could you please provide the code of Pacman envieronment mentioned in the paper? Some details about this environment are not very clear.

The video on your Google sites is very interesting. Could you please tell me how to save the replay in the football env?