Learning Summarised Messaging Through Mediated Differentiable Inter-Agent Learning

Abstract In recent years, notable research has been done in the area of communication in multi-agent systems. When agents have a partial view of the environment, communication becomes essential for collaboration. We propose a Deep Q-Learning based multi-agent communication approach: Mediated Differentiable Inter-Agent Learning (M-DIAL), where messages produced by individual agents are sent to a mediator that encodes all the messages into a global embedding. The mediator essentially summarises the crux of the messages it receives into a single global message that is then broadcasted to all the participating agents. The proposed technique allows the agents to receive only essential abstracted information and also reduces the overall bandwidth required for communication. We analyze and evaluate the performance of our approach over several collaborative multi-agent environments.

Authors

• Sharan Gopal
• Rishabh Mathur
• Shaunak Deshwal
• Anil Singh Parihar

Keywords

No keywords are indexed for this paper.