Learning multi-agent coordination through connectivity-driven communication

Abstract In artificial multi-agent systems, the ability to learn collaborative policies is predicated upon agents’ communication skills: they must be able encode information received from environment and how share it with other agents as required by task at hand. We present a deep reinforcement learning approach, Connectivity Driven Communication (CDC), that facilitates emergence of behaviour only through experience. The are modelled nodes weighted graph whose state-dependent edges pair-wise messages can exchanged. introduce graph-dependent attention mechanisms controls incoming weighted. This mechanism takes into full account current state system represented graph, builds diffusion process captures flows on graph. topology not assumed known priori, but depends dynamically observations, learnt concurrently policy in an end-to-end fashion. Our empirical results show CDC effective over-perform competing algorithms cooperative navigation tasks.