Research map / Multi-agent systems and games / Cooperative multi-agent RL
CoDe: Communication Delay-Tolerant Multi-Agent Collaboration via Dual Alignment of Intent and Timeliness
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 CoDe, a delay-tolerant multi-agent communication framework that learns future-action-inferred intent messages and fuses asynchronous messages via intent and timeliness attention alignment.
Problem
Multi-agent reinforcement learning with asynchronous communication, where channel delays cause agents to receive messages sent at different time points, leading to cognitive bias and collaboration breakdown.
Approach
Agents send a learned intent representation trained via future action inference (inference loss) and temporal continuity (continuity loss); receivers fuse messages with attention over intents plus an exponential temporal discount on delayed messages (dual alignment of intent and timeliness).
Key results
- Under fixed delay 3 on SMAC map 2s3z_vs_1ul, CoDe reaches 89.3 ± 3.3 versus 43.1 ± 6.9 for the best baseline TarMAC.
- Under time-varying delay N(5,2) on SMAC 1c4z_vs_1ul, CoDe reaches 67.4 ± 4.6 versus 49 ± 7.8 for TarMAC.
- Ablation on 1o_10b_vs_1r shows CoDe at 62.1 vs 35.6 without intent alignment and 33.8 without dual alignment.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add intent alignment (ablation) | CoDe · SMAC 1o_10b_vs_1r, time-varying delay N(3,2) | win rate: 35.6 -> 62.1 | Helped |
| add timeliness alignment (ablation) | CoDe · SMAC 1o_10b_vs_1r, time-varying delay N(3,2) | win rate: 58.5 -> 62.1 | Helped |
| add dual alignment (ablation) | CoDe · SMAC 1o_10b_vs_1r, time-varying delay N(3,2) | win rate: 33.8 -> 62.1 | Helped |
| add the inference loss (ablation) | CoDe intent learning · SMAC delayed environments | performance | Helped |
| add the continuity loss (ablation) | CoDe intent learning · SMAC delayed environments | performance | Helped |
| add both inference and continuity losses (ablation) | CoDe intent learning · SMAC delayed environments | performance | Helped |
| remove intent regularization losses (ablation) | CoDe intent learning · SMAC delay-free environments | performance | No clear effect |
Surprising findings
- Conventional communication MARL methods can perform worse than no communication under asynchronous delays.
- Sparse-communication methods such as NDQ and TarMAC show some delay robustness.
Limitations
- Limited by the length of future action prediction.
- Struggles with more severe delays.
- Intent alignment and timeliness alignment assume zero delay during training, so training and testing differ.
Experimental setup
Datasets: SMAC GRF Hallway
Benchmarks: SMAC GRF Hallway
Metrics: win rate team scoring rate episode reward
Baselines: QMIX TarMAC NDQ TMC MAIC CACOM T2MAC
Builds on: QMIX NDQ TarMAC MAIC T2MAC
Closest papers by meaning
- Narrowing the Collaboration Gap, Probably (NeurIPS 2026 · similarity 0.87)
- T2MAC: Targeted and Trusted Multi-Agent Communication through Selective Engagement and Evidence-Driven Integration (AAAI 2024 · similarity 0.87)
- Multi-Agent Coordination via Multi-Level Communication (NeurIPS 2024 · similarity 0.87)
- COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.86)
- Semantics and Spatiality of Emergent Communication (NeurIPS 2024 · similarity 0.86)
- RADAR: Redundancy-Aware Diffusion for Multi-Agent Communication Structure Generation (ICML 2026 · similarity 0.85)
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.