This repository provide a basic framework for training agents to communicate using RLlib. The agents communicate via a differentiable channel that allows us to backpropagate through the communication channel. Thereby, the parameters of the sender agent can be optimised to maximise the reward generated by the receiver agent's actions in the environment.
dylancope / rllib_emecom Goto Github PK
View Code? Open in Web Editor NEWEmergent Communication in RLlib
License: MIT License