New Framework Boosts Multi-Agent Robustness in Noisy Environments

Cooperative multi-agent reinforcement learning (MARL) systems often struggle with maintaining effective coordination when agents receive noisy observations. While individual observation disturbances might seem independent, their collective impact on decision-making can become structured, especially within groups of agents with strong task dependencies. This phenomenon, termed "structured noise effects," means noise-induced decision deviations are locally correlated but globally heterogeneous. To address this, researchers propose SIGMA (Structured Noise-Effect-Aware Grouped Multi-Agent Aggregation). This hierarchical framework is designed to leverage these cooperation structures to learn more robust representations. SIGMA first dynamically groups agents based on density and then performs intra-group consensus aggregation to filter out agent-specific noise while retaining shared, task-relevant information. Following intra-group processing, an inter-group attention mechanism adaptively integrates information across different groups. This ensures global coordination is maintained while accommodating the unique contributions and characteristics of each group. Experimental results in StarCraft II environments with noisy observations confirm the existence of structured noise effects and demonstrate SIGMA's consistent improvement in robustness without sacrificing performance in ideal conditions.