Research map / Multi-agent systems and games / Cooperative multi-agent RL
Learning to Communicate Through Implicit Communication Channels
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 the Implicit Channel Protocol (ICP), which encodes messages into designated "scouting" actions that act as an implicit communication channel, and trains agents with a random initial or delayed information map, reaching 24.91/25 in Hanabi.
Problem
Communication in collaborative multi-agent RL when explicit messaging channels are unavailable; existing implicit communication relies on theory-of-mind inference, which is computationally costly and inaccurate in complex tasks.
Approach
Designate a subset of actions (scouting actions) with no or uniform effect on environment dynamics as an implicit channel, use a shared centralized mapping from messages to these actions so receivers decode intent directly without ToM inference, and train with either a randomly initialized map (DIAL-style Gumbel-Softmax gradients) or a delayed map (pre-train on explicit channel, then transfer).
Key results
- On Hanabi (4 players), ICN-RGMComm-DM-hat-mapping achieves an average score of 24.91/25 versus SAD's 23.81 and average human 17.
- On Guessing Numbers (N=3, l=11), ICN-DIAL-RM averages about 8.63 episode steps versus about 12.4 for VDN on-policy and off-policy.
- On Revealing Goals, ICN-DIAL-RM achieves an average score 2.17 times that of the baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ICP with hat mapping instead of the SAD baseline | Hanabi 4-player full version · Hanabi | average score: 23.81 -> 24.91 | Helped |
| Use ICP instead of the SAD baseline | Hanabi 4-player full version · Hanabi | win rate: 41.45% -> 91% | Helped |
| Use ICP instead of ToM-based methods | Hanabi 4-player full version · Hanabi | average score: 19.13 -> 24.91 | Helped |
| Use ICP instead of average human play | Hanabi 4-player full version · Hanabi | average score: 17 -> 24.91 | Helped |
| Use the random-initial-map ICP strategy instead of VDN baselines | Guessing Numbers · Guessing Numbers | average episode length: around 12.4 -> approximately 8.63 | Helped |
| Use the random-initial-map ICP strategy instead of the baselines | Revealing Goals · Revealing Goals | average score: 2.17 times | Helped |
| Use hat mapping rather than the other tested ICP variants (ablation) | Hanabi ICP · Hanabi 4-player full version | average score | Helped |
| Shuffle the embedding while keeping the information strategy fixed, then fine-tune the action policy (ablation) | ICN-RGMComm-hat-mapping policy · Hanabi | performance | Mixed |
Surprising findings
- Randomly shuffling the scouting-action embeddings while keeping the information strategy fixed sometimes improves performance, suggesting environment-based information also contributes.
Limitations
- Requires the environment to contain natural scouting actions with no or uniform effect on dynamics (e.g., Hanabi hints); extension to environments lacking them is future work.
- Hanabi results rely on best runs selected from more than three seeds, and score differences vs. SAD are small due to the 25-point ceiling.
Experimental setup
Datasets: Guessing Numbers (self-designed environment) Revealing Goals (self-designed environment) Hanabi
Benchmarks: Guessing Numbers Revealing Goals Hanabi
Metrics: average score average episode length win rate
Baselines: VDN-on-policy VDN-off-policy DIAL-Cheat SAD WTFWThat+search Theory of mind-based methods
Builds on: DIAL VDN RGMComm SAD Hat guessing game hat mapping
Closest papers by meaning
- Semantics and Spatiality of Emergent Communication (NeurIPS 2024 · similarity 0.87)
- Efficient Communication in Multi-Agent Reinforcement Learning with Implicit Consensus Generation (AAAI 2025 · similarity 0.86)
- Implicit Intelligence - Evaluating Agents on What Users Don’t Say (ICML 2026 · similarity 0.86)
- Narrowing the Collaboration Gap, Probably (NeurIPS 2026 · similarity 0.86)
- Mental-Models for Multi-Agent Systems (NeurIPS 2026 · similarity 0.86)
- Thought Communication in Multiagent Collaboration (NeurIPS 2025 · 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.