Logo image
Back
Encoding Goals as Graphs: AAMAS '26: Proceedings of the 25th International Conference on Autonomous Agents and Multiagent Systems
Conference proceeding

Encoding Goals as Graphs: AAMAS '26: Proceedings of the 25th International Conference on Autonomous Agents and Multiagent Systems

Alessandro Amato, Raffaele Galliera, K. Brent Venable and Niranjan Suri
Proc. of the 25th International Conference on Autonomous Agents and Multiagent Systems, pp.3513-3515
ACM Conferences
AAMAS 2026: Autonomous Agents and Multiagent Systems, 25th (Paphos Cyprus, 05/25/2026–05/29/2026)
05/25/2026
DOI: https://doi.org/10.65109/TDTU5449
Web of Science ID: WOS:001880341000471

Metrics

4 Record Views

Abstract

Multi-agent reinforcement learning Graph neural networks Goalconditioned learning Graph embeddings Representation learning Cognitive or Behavioral Models Electronic Neural Network Higher Education
Many cooperative multi-agent tasks are naturally defined by graph-structured objectives, where agents must collectively achieve a desired relational configuration or satisfy a set of constraints. However, current goal-conditioned multi-agent reinforcement learning (MARL) methods rarely leverage such symbolic structure to guide learning. To address this challenge, we propose Graph Embeddings for Multi-Agent Coordination (GEMA), which augments any cooperative learner with a State-Graph Encoder (SGE). The SGE is pre-trained contrastively to embed state and goal graphs in a shared metric space. At run time, each agent constructs the state graph, queries the SGE, and computes a similarity score to the goal embedding. This similarity serves as an intrinsic reward, providing dense feedback on task progress, and is also incorporated into each agent's observation. Experiments on cooperative navigation, load balancing, and the StarCraft Multi-Agent Challenge (v2) show that GEMA accelerates convergence and improves team returns.

Details

Logo image