Abbreviated Weighted Graph in Multi-Agent Reinforcement Learning

The cooperation phenomenon in real life can often be viewed as an interaction between multiple agents. The exchange of unstructured information between agents is usually abstracted as the transmission of node information in a graph. Various approaches have been proposed to build graph models and fuse the information of agents in their observation range. However, it is inefficient to directly apply the related algorithms of graph neural networks to integrate the information of each agent, which can not effectively use the observation features of the dominant agent in cooperation, and it may also be overly focused on the weight of the edges. In this paper, we propose AWGmix, which is an abbreviated weighted graph information-enhanced algorithm for multi-agent reinforcement learning. Specifically, we propose a simple and convenient method to calculate the weight of edges between graph nodes modeled from agent connections, and design an attribution module based on attention mechanism to find the dominant agents and enrich the representation of observations. Experimental results demonstrate the superiority and effectiveness of our proposed method on Starcraft II micromanagement benchmark tasks.