EDA3-X: Distributed Attentional Actor Architecture for Interpretability of Coordinated Behaviors in Multi-Agent Systems

Abstract

In this paper, we propose an enhanced version of the distributed attentional actor architecture (eDA3-X) for model-free reinforcement learning. This architecture is designed to facilitate the interpretability of learned coordinated behaviors in multi-agent systems through the use of a saliency vector that captures partial observations of the environment. Our proposed method, in principle, can be integrated with any deep reinforcement learning method, as indicated by X, and can help us identify the information in input data that individual agents attend to during and after training. We then validated eDA3-X through experiments in the object collection game. We also analyzed the relationship between cooperative behaviors and three types of attention heatmaps (standard, positional, and class attentions), which provided insight into the information that the agents consider crucial when making decisions. In addition, we investigated how attention is developed by an agent through training experiences. Our experiments indicate that our approach offers a promising solution for understanding coordinated behaviors in multi-agent reinforcement learning.