Research map / Multi-agent systems and games / Cooperative multi-agent RL
Efficient Communication in Multi-Agent Reinforcement Learning with Implicit Consensus Generation
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 COCOM, a MARL framework that combines implicit consensus (learned via a VAE-based global consensus generator aligned with local consensus) with gated, mutual-information-regularized messages to cut communication overhead while keeping cooperative performance.
Problem
Multi-agent cooperation under partial observability faces a trade-off: explicit communication reduces teammate uncertainty but has high bandwidth costs, while implicit consensus is cheap but degrades when local observations are very limited.
Approach
Agents form a local consensus from their observations (aligned to a training-only global consensus via a VAE), then generate messages conditioned on that consensus, using mutual-information objectives to make messages non-redundant and useful and a learned gate to decide when to send them.
Key results
- On Hallway, COCOM achieves high win rates while communicating only at critical timesteps, whereas Full-Comm is overwhelmed by redundant messages and TMC over-compresses and fails the task.
- COCOM attains the lowest communication overhead among compared methods on 2c_vs_64zg and 10m_vs_11m in SMAC while maintaining superior win rates.
- Integrating COCOM with VDN and QMIX significantly improves both base algorithms on GRF academy 3 vs 1 with keeper.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine implicit consensus with gated explicit communication (COCOM) | Multi-agent cooperative policies · Hallway | test win rate | Helped |
| Combine implicit consensus with gated explicit communication (COCOM) | Multi-agent cooperative policies · SMAC | test win rate | Helped |
| Use COCOM's adaptive gating to reduce message transmission | COCOM · SMAC (2c vs 64zg, 10m vs 11m) | communication overhead | Helped |
| Combine COCOM with value decomposition | VDN and QMIX · GRF academy 3 vs 1 with keeper | test win rate | Helped |
Surprising findings
- Full-Comm, which sends unfiltered messages, degrades performance on SMAC due to redundant information.
- Shutting down communication entirely can cause training instability for some filtering methods, such as TMC on Hallway.
Limitations
- Messages are broadcast to all teammates rather than customized per recipient.
- Consensus learned from local observations can struggle when observation discrepancies among agents are large.
Experimental setup
Datasets: Hallway SMAC Google Research Football
Benchmarks: Hallway SMAC (2s3z, 3s5z, 1c3s5z, 10m_vs_11m, 2c_vs_64zg, MMM2) Google Research Football academy 3 vs 1 with keeper
Metrics: test win rate communication overhead (fraction of communicating agent pair
Baselines: QMIX VDN COLA Full-Comm TarMAC NDQ MAIC TMC
Builds on: QMIX COLA CLUB NDQ MAIC
Closest papers by meaning
- Multi-Agent Incentive Communication via Decentralized Teammate Modeling (AAAI 2022 · similarity 0.91)
- Multi-Agent Reinforcement Learning with Communication-Constrained Priors (NeurIPS 2025 · similarity 0.90)
- Learning Multi-Agent Communication from Graph Modeling Perspective (ICLR 2024 · similarity 0.89)
- Multi-Agent Coordination via Multi-Level Communication (NeurIPS 2024 · similarity 0.89)
- Exponential Topology-enabled Scalable Communication in Multi-agent Reinforcement Learning (ICLR 2025 · similarity 0.88)
- Bridging Training and Execution via Dynamic Directed Graph-Based Communication in Cooperative Multi-Agent Systems (AAAI 2025 · similarity 0.88)
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.