SIGMA is a hierarchical collaboration framework that learns robust representations for multi-agent reinforcement learning under noisy observations. It first employs density‑based grouping to partition agents into adaptive local structures, then performs intra‑group consensus aggregation to preserve shared task‑relevant information while smoothing individual representation deviations. An inter‑group attention module subsequently integrates information across groups adaptively, maintaining global coordination and accommodating heterogeneous contributions. Experiments on noisy‑observation tasks in StarCraft II confirm the presence of structured noise effects and demonstrate that SIGMA consistently improves robustness under observation noise while retaining competitive performance in noise‑free settings.
Blogger's Review: By explicitly exploiting the structure‑dependent nature of observation noise, the approach offers a practical avenue for enhancing MARL stability in real‑world noisy environments.