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

2026-04-30 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, short

Summary

Valentin Cuzin-Rambaud, Laetitia Matignon, and Maxime Morge from LIRIS, UCBL, have published a survey on Multi-Agent Deep Reinforcement Learning (MARL) that incorporates Graph Neural Network (GNN)-based communication, submitted on April 28, 2026. This work addresses the growing research in MARL where agents use GNNs to learn communication, thereby enhancing their internal representations through shared information. The authors note a lack of a clear framework for classifying these GNN-based communication approaches within MARL. Their survey proposes a generalized GNN-based communication process to clarify the underlying concepts and make them more accessible. This paper was presented at the Rencontres des Jeunes Chercheurs en Intelligence Artificielle (RJCIA) during the Plate-Forme Intelligence Artificielle (PFIA) in June 2026, in Arras, France.

Key takeaway

For research scientists developing multi-agent systems, understanding the proposed generalized GNN-based communication framework can help structure your approach to inter-agent communication. You should consider how GNNs can be integrated to facilitate information exchange and improve coordination, potentially leading to more robust and efficient MARL solutions.

Key insights

GNNs facilitate multi-agent coordination by enabling learned communication and information sharing.

Principles

• Communication improves multi-agent coordination.
• GNNs can learn effective communication protocols.

Method

The survey proposes a generalized GNN-based communication process to structure and clarify existing MARL approaches that use GNNs for inter-agent communication.

In practice

• Apply GNNs for agent communication.
• Enhance agent representations via shared data.

Topics

• Multi-Agent Reinforcement Learning
• Graph Neural Networks
• Agent Communication
• Multiagent Systems
• Reinforcement Learning

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Student

Related on AIssential

• Learning to Communicate: Toward End-to-End Optimization of Multi-Agent Language Systems — DiffMAS optimizes multi-agent systems by making latent communication a jointly learnable component.
• Interview with Xinwei Song: strategic interactions in networked multi-agent systems — Strategic interactions in networked multi-agent systems can be managed through mechanism design and reputation-aware reinforcement learning.
• Learning Altruistic Collaboration in Heterogeneous Multi-Team Systems — A GNN policy extends Hamilton's rule for altruistic robot allocation in heterogeneous multi-team systems, solving NP-hard problems scalably.
• LLM-Guided Communication for Cooperative Multi-Agent Reinforcement Learning — LLMs can iteratively refine communication protocols in MARL to improve agents' shared state awareness.
• Coordination Graphs for Constrained Multi-Agent Reinforcement Learning — The CG-CMARL framework efficiently scales multi-agent reinforcement learning by combining coordination graphs and Lagrangian duality for constraint handling.
• Complex & Intelligent Systems, Volume 12, Issue 1, January 2026 — The volume showcases diverse AI applications, emphasizing advanced neural networks, optimization, and multimodal data fusion.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.