Interaction-Breaking Adversarial Learning Framework for Robust Multi-Agent Reinforcement Learning

Cooperation is central to multi-agent reinforcement learning (MARL), yet learned coordination can be fragile when external perturbations disrupt inter-agent interactions. Prior robust MARL methods have primarily considered value-oriented attacks, leaving a gap in robustness when interaction structures themselves are corrupted. In this paper, we propose an interaction-breaking adversarial learning (IBAL) framework that takes an information-theoretic view to construct attacks that impede coordination by perturbing agents’ observations and actions, and trains agents to perform reliably under such disruptions. Empirically, our approach improves robustness over existing robust MARL baselines across diverse attack settings and yields stronger performance even under agent-missing scenarios. Our code is available at https://sunwoolee0504.github.io/IBAL.

When several AI agents work together, such as robots, game units, or autonomous drones, their success often depends on how well they coordinate with one another. However, this coordination can break down when some agents cannot properly observe, influence, or respond to their teammates. In this work, we study this kind of failure and ask how multi-agent AI systems can remain reliable when their interactions are disrupted. We propose Interaction-Breaking Adversarial Learning, or IBAL, a training framework that deliberately creates difficult situations where agents lose important information about one another or have their actions disturbed. Instead of only attacking agents to make them choose bad actions, IBAL focuses on breaking the connections that support teamwork. By repeatedly training under these disrupted conditions, agents learn to reorganize their behavior and continue cooperating even when coordination partially fails. Across several cooperative benchmark tasks, IBAL makes agents more robust than previous methods under different attacks and also when some teammates are missing or weakened. This suggests that preparing AI teams for interaction failures can lead to more dependable multi-agent systems in real-world settings.