Interpretable Multi-Agent Communication via Information Gating

Stav Belogolovsky, Eran Iceland, Itay Naeh, Ariel Barel, Shie Mannor

Keywords: interpretability, communication, MARL

Abstract: Multi-Agent Reinforcement Learning (MARL) holds significant promise for complex coordination tasks, where communication is often vital. However, emergent communication in MARL is typically opaque and unintelligible to humans, and forcing human-like language can impede learning or performance. Furthermore, agents struggle when processing high-dimensional unstructured communication. Addressing these critical challenges, we propose a novel framework enabling agents to learn effective communication while ensuring its interpretability. Our core idea is to make the subject of communication transparent, rather than the message content itself. This is achieved by training agents to learn a policy that selectively gates the information flow from their observation to the communication channel. By adopting an object-oriented perspective and using a text-to-mask model that maps terms to observation features, agents learn to select and communicate only relevant information. This approach provides enhanced interpretability by revealing message context, and mitigates information overload through selective masking. We introduce this comprehensive framework, demonstrating its effectiveness and robustness across multi-agent tasks that require communication. We analyze emergent communication protocols and their resulting interpretability and release our code and environments to support further research.

Submission Number: 10