mansicer / maic Goto Github PK

Why maximize Mutual Information between a_{j} and latent variable z_{ij}？

Thank you for your awesome work!

I was trying to reproduce Fig. 5 in your AAAI paper, and I used the default parameters for the MAIC algorithm on map 27m_vs_30m with a tiny change (batch_size=16, buffer_size=16 due to my GPU resource limitation, seed=0; seed=1), and I tried two sets of learning rates (lr=0.0005, critic_lr=0.0005 as in the default configs and lr=0.00025, critic_lr=0.00025 since I reduce the batch size).

But after 2M environment timesteps, I still got the winning rate of 0 for both training and testing for all the settings I mentioned above. The return median is somewhere between 3 and 5. So I was wondering if there is some discrepancy between the default parameters in this repository and the ones used in Fig. 5 in your paper. It would be great if you could kindly release those settings.

Thank you very much!

Hello, I have installed the Python environment according to the requirements.txt file, but when I run the following command:
python src/main.py --config=qmix --env-config=foraging

The following error occurred

How can I solve this mistake? is there any other environment to install?
Looking forward to your reply.

你好，最近在研读您的文章，有个问题想请教一下。文中提到：
exchange local observation /embedding

This informative communication scheme can reduce uncertainty for individual local decision-
making and induce implicit coordination. However, such a communication mechanism enlarges agents’ local policy
spaces and increases learning complexity,.
我想问下这里的 enlarges agents’ local policy spaces and increases learning complexity, 的 local policy spaces是指扩大local observation吗？

In q_learner.py line 112:
td_error = (chosen_action_qvals - targets.detach())
However, in the reference, I see, the td_error should be
td_error = (targets.detach() - chosen_action_qvals)

I think if you refer to the GitHub Repository Python MARL framework? Cause in this Repository, the td_error is also:
td_error = (chosen_action_qvals - targets.detach())

In this paper, Section in "Experimental Results"

what is 25%-75% deviation?
Can you give a simple example？

Hello, it's said in the paper that the experiment on SMAC was tested on "all 13 maps",(Figure 6).
However, $ python -m smac.bin.map_list, would return 38 maps.
Could you please let me know, which 13 maps did you used for the experiments in the paper.

I get a error information when I try to run this code.

Traceback (most recent calls WITHOUT Sacred internals):
File "src/main.py", line 35, in my_main
run(_run, config, _log)
File "/home/liyurui/projects/MAIC/src/run.py", line 64, in run
run_sequential(args=args, logger=logger)
File "/home/liyurui/projects/MAIC/src/run.py", line 258, in run_sequential
learner.train(episode_sample, runner.t_env, episode)
File "/home/liyurui/projects/MAIC/src/learners/maic_learner.py", line 60, in train
mixer=self.target_mixer if self.supervised_loss_with_mixer else None,
AttributeError: 'MAICLearner' object has no attribute 'supervised_loss_with_mixer'

您好，请问您 maic_agent.py 代码文件中的 key 和 query 是不是弄反了？你把 trajectory 做成了 key，队友建模采样出来的 latent 作为 query，好像跟您原文不太一样？

Thank you for your wonderful work, and I really enjoyed reading it.

Just a quick question, could you please kindly provide a pointer to the superhard maps(MMM3, 3c_vs_100zg, etc.) that you used for the paper?

Thanks a lot

When I change the configuration in foraging.yaml like this:

env_args:
  field_size: 20
  players: 10
  max_food: 6
  force_coop: False
  partially_observe: True
  sight: 3
  is_print: False
  need_render: False

There will be an error report about gym.error.UnregisteredEnv.

So I wonder whether the environment configuration in foraging.yaml can be modified?
Looking foward to your reply.Thank you very much.

why in maic_agent.py, returns['entropy loss'] is not used in maic_learner?