Research map / Multi-agent systems and games / Cooperative multi-agent RL
T2MAC: Targeted and Trusted Multi-Agent Communication through Selective Engagement and Evidence-Driven Integration
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 T2MAC, a multi-agent RL method that learns targeted (selective partner/timing) communication and evidence-level message integration via Dirichlet/subjective logic and Dempster-Shafer combination, evaluated on Hallway, MPE and SMAC.
Problem
Broadcast communication in cooperative multi-agent RL is inefficient and redundant, and existing methods integrate received messages with basic aggregation or black-box representation learning, leading to uncertain integration.
Approach
Agents extract evidence (Dirichlet parameters via subjective logic) to make decisions and craft tailored messages per teammate; a selector network trained on pseudo-labels from uncertainty reduction chooses when and with whom to communicate; received evidence is fused with Dempster-Shafer combination rather than a learned black-box fusion.
Key results
- T2MAC reports 37.2% performance improvement from communication, 56.0% communication rate and 66.4% communication efficiency on SMAC, versus 30.2% for MASIA (100% comm rate) in the reported table.
- T2MAC matches MASIA on the Hallway task under harder conditions, while other communication baselines falter as difficulty increases.
- Ablation on 1o_10b_vs_1r shows that the Dirichlet policy alone (T2MAC without communication) improves over QMIX, and evidence-driven integration and selective engagement each add performance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use T2MAC instead of MASIA | Communication method · SMAC | Comm efficiency: 30.2% -> 66.4% | Helped |
| Use T2MAC instead of MASIA | Communication method · SMAC | Performance improvement: 30.2% -> 37.2% | Helped |
| Use T2MAC instead of MASIA | Communication method · SMAC | Comm rate: 100.0% -> 56.0% | Helped |
| Replace the categorical distribution with a Dirichlet distribution (ablation) | T2MAC Q-value network · 1o_10b_vs_1r | Win rate | Helped |
| Add communication with evidence-driven information exchange and integration (ablation) | T2MAC without communication · 1o_10b_vs_1r | Win rate | Helped |
| Add selective engagement instead of communicating continuously with all agents (ablation) | T2MAC Fullcomm · 1o_10b_vs_1r | Win rate | Helped |
| Use evidence-driven integration instead of summation-based integration (ablation) | T2MAC · 1o_10b_vs_1r | Win rate | Helped |
| Use evidence-driven integration instead of black-box TarMAC integration (ablation) | T2MAC · 1o_10b_vs_1r | Win rate | Helped |
Surprising findings
- All communication methods significantly outperform non-communication methods, suggesting the chosen environments demand communication.
Limitations
- Communication selection relies on a constant threshold to create pseudo-labels, which is not analyzed for sensitivity in the text.
- Evaluation is limited to the Hallway, MPE and SMAC environments with a small set of baselines.
Experimental setup
Datasets: Hallway MPE Cooperative Navigation MPE Predator Prey SMAC
Benchmarks: Hallway MPE Cooperative Navigation MPE Predator Prey SMAC
Metrics: win rate performance improvement communication rate communication efficiency
Baselines: QMIX DOP TarMAC MAIC SMS MASIA CommNet MAPPO
Builds on: QMIX TarMAC MAIC MASIA Subjective Logic
Closest papers by meaning
- Systematic Failures in Collective Reasoning under Distributed Information in Multi-Agent LLMs (ICML 2026 · similarity 0.88)
- Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems (ICLR 2025 · similarity 0.87)
- Narrowing the Collaboration Gap, Probably (NeurIPS 2026 · similarity 0.87)
- COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.87)
- CoDe: Communication Delay-Tolerant Multi-Agent Collaboration via Dual Alignment of Intent and Timeliness (AAAI 2025 · similarity 0.87)
- Efficient Communication in Multi-Agent Reinforcement Learning with Implicit Consensus Generation (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.