Research map / Multi-agent systems and games / Cooperative multi-agent RL
HYGMA: Hypergraph Coordination Networks with Dynamic Grouping for Multi-Agent Reinforcement Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HYGMA, which uses dynamic spectral clustering on agent state histories to build hypergraphs and processes them with attention-based hypergraph convolution, plugged into QMIX-style value-based and actor-critic policy-based MARL.
Problem
Cooperative multi-agent RL needs agent organization and information exchange that adapt dynamically, while fixed pairwise or static grouping structures fail to capture evolving higher-order coordination.
Approach
Builds hyperedges from dynamic spectral clustering of agents' state histories (k-NN similarity graph, silhouette-selected k, stability-thresholded regrouping) and applies attention-weighted hypergraph convolution to produce group-aware features fed into Q-functions (QMIX) or policies/critics.
Key results
- On Traffic Junction 7x7, HYGMA reaches 99.7% success (272 steps) versus MAGIC's 99.9% (440 steps), i.e. comparable final success with faster convergence.
- On GRF, HYGMA achieves 97.7% success rate versus MAGIC's 98.2%, with fewer steps taken (30.26 vs 34.30).
- Reported HGCN training overhead is about 36% over Ft-QMIX wall-clock time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HYGMA instead of MAGIC | Traffic Junction · 7×7, Nmax=5, parr ive=0.3 | convergence steps: 440 ± 64 -> 272 ± 41 | Helped |
| Use HYGMA instead of MAGIC | Traffic Junction · 7×7, Nmax=5, parr ive=0.3 | success rate: 99.9 ± 0.1% -> 99.7 ± 0.1% | Hurt |
| Use HYGMA instead of MAGIC | Traffic Junction · 14×14, Nmax=10, parr ive=0.2 | convergence steps: 819 ± 85 -> 569 ± 34 | Helped |
| Use HYGMA instead of MAGIC | Traffic Junction · 14×14, Nmax=10, parr ive=0.2 | success rate: 99.9 ± 0.1% -> 99.2 ± 0.1% | Hurt |
| Use HYGMA instead of MAGIC | Google Research Football · Google Research Football | steps taken: 34.30± 1.34 -> 30.26± 0.24 | Helped |
| Use HYGMA instead of MAGIC | Google Research Football · Google Research Football | success rate: 98.2 ± 1.0% -> 97.7± 0.7% | Hurt |
| Replace standard GCN with HGCN while retaining dynamic grouping (ablation) | HYGMA · SMAC 5m_vs_6m | win rate | Helped |
| Replace a fixed single group with dynamic grouping and hypergraph structure (ablation) | HYGMA · SMAC 5m_vs_6m | win rate | Helped |
Surprising findings
- On GRF and Traffic Junction, HYGMA does not beat MAGIC in final success rate; MAGIC is marginally higher.
Limitations
- Theoretical guarantees (log k approximation, convergence, quality preservation) rely on loose or unproven steps in the appendix.
- Final success rates on Traffic Junction and GRF are below MAGIC.
- Training overhead of about 36% relative to the QMIX baseline.
Experimental setup
Datasets: SMAC Predator-Prey Traffic Junction Google Research Football
Benchmarks: SMAC 3s_vs_5z SMAC 5m_vs_6m SMAC 3s5z_vs_3s6z Predator-Prey Traffic Junction Google Research Football
Metrics: win rate success rate convergence steps steps taken
Baselines: Ft-QMIX QPLEX VAST MAPPO GoMARL RIIT MAGIC CommNet GA-Comm I3CNet TarMAC-IC3Net IC3Net
Builds on: QMIX VAST GoMARL MAGIC
Closest papers by meaning
- HyperMARL: Adaptive Hypernetworks for Multi-Agent RL (NeurIPS 2025 · similarity 0.87)
- Bridging Training and Execution via Dynamic Directed Graph-Based Communication in Cooperative Multi-Agent Systems (AAAI 2025 · similarity 0.86)
- Graph Diffusion for Robust Multi-Agent Coordination (ICML 2025 · similarity 0.86)
- Learning Multi-Agent Communication from Graph Modeling Perspective (ICLR 2024 · similarity 0.86)
- HGAP: Boosting Permutation Invariant and Permutation Equivariant in Multi-Agent Reinforcement Learning via Graph Attention Network (ICML 2024 · similarity 0.86)
- Towards Efficient Collaboration via Graph Modeling in Reinforcement Learning (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.