I am trying to use keras-rl but in a multi-agent environment. So I found this github issue of keras-rl with an idea using shared environment for all agents. Unfortunately, I haven't managed to get it working. It seems that using a gym environment in such context is possible in general, I just haven't managed to get it working with keras-rl. Anybody that know how to do it?
1 Answers
I put together something that works for sequential games based on one of the replies in the issue you linked, and it mostly seems to work: https://github.com/velochy/rl-bargaining/blob/master/interleaved.py
It should be rather straightforward to adapt to simultaneous games by adjusting the forward and backward functions to split the action space appropriately and call all the sub-agents.
Related