openai/multiagent-particle-envs

Communication Signal in simple_reference

isaeed3 opened this issue · 6 comments

Hi,

Can anyone explain why is the communication signal of size 10 in simple_reference. It is the output of actor NN along with physical action of size 5 but i couldn't figure out why is it chosen to be of size 10? Is it chosen by hit and trial or is there any particular reason of it being of size 10?

Hi, may we discussed more about this? I'm currently using this env for my works. I'm wondering the details about the communication information. I can't figure out the content included in the communication information.

The communication vector is actually incomprehensible and the communication protocol emerges purely from training to maximise rewards.
So the size 10 seems to be chosen by hit and trial.

Thank you for your reply. Therefore, do u mean that the communication information were learned during the learning process. These is no presumed details stipulated by us? But when I checked the core data, I found the communication information were assigned to the same as the actions of leader adversary. Any explanation for this?

It's been a while since I had gone through the code. However, as far as I remember, the input to the NN is provided by us but its output is used as the communication signal that is of course based on the learning procedure.

Thank you again. It sounds making sense. However, I got a new question from that, the dimension of the communication information should be the same as the dimension of the action space of the leader adversary at this time, why should we choose a proper dimension for the communication channel by hit and trial?

If you see other environments, the dimension of communication signal is different in all environments. That's why I posted my question here as to how the dimension is determined