Learning Attentional Communication with a Common Network for Multiagent Reinforcement Learning.

Abstract

For multiagent communication and cooperation tasks in partially observable environments, most of the existing works only use the information contained in hidden layers of a network at the current moment, limiting the source of information. In this paper, we propose a novel algorithm named multiagent attentional communication with the common network (MAACCN), which adds a consensus information module to expand the source of communication information. We regard the best-performing overall network in the historical moment for agents as the common network, and we extract consensus knowledge by leveraging such a network. Especially, we combine current observation information with the consensus knowledge to infer more effective information as input for decision-making through the attention mechanism. Experiments conducted on the StarCraft multiagent challenge (SMAC) demonstrate the effectiveness of MAACCN in comparison to a set of baselines and also reveal that MAACCN can improve performance by more than 20% in a super hard scenario especially.