Multi Agent Reinforcement Learning with Graph Policy Gradients

We are pleased to announce we have open sourced our code for very large scale multi-agent reinforcement learning. Our core idea involves parametrization of policies with GCNs for multi-agent problems instead of the traditional FCNs. We utilize DGL to implement the policies for the multiple agents.

Our paper was presented at the Conference of Robot Learning 2019 as an oral paper.
Code, details about the paper and the talk can be found here:


Great job! Thank you for sharing.