A Survey of Multi-Agent Deep Reinforcement Learning with Graph Neural Network-Based Communication

In multi-agent reinforcement learning (MARL), agents must coordinate actions to achieve shared objectives, but effective coordination often requires robust communication. A new survey by Valentin Cuzin-Rambaud, Laetitia Matignon, and Maxime Morge from LIRIS, UCBL explores how graph neural networks (GNNs) can enhance this communication. The paper, submitted to arXiv on April 28, 2026, proposes a generalized GNN-based communication process where agents use interaction graphs to exchange information, enriching their internal representations. This approach helps agents learn to share relevant data, improving convergence and task performance in complex environments.

The survey addresses a critical gap: the lack of explicit structure to classify MARL methods with GNN-based communication. By analyzing recent works, the authors provide a framework that distinguishes different approaches, making underlying concepts more obvious and accessible. The research spans machine learning, AI, and multiagent systems, with practical implications for robotics, autonomous vehicles, and distributed control. The paper was presented at the Rencontres des Jeunes Chercheurs en Intelligence Artificielle (RJCIA) during the PFIA 2026 conference in Arras, France. This work offers a valuable roadmap for researchers looking to implement or advance GNN-driven coordination in multi-agent systems.