Research map / Multi-agent systems and games / Cooperative multi-agent RL
Expressive Multi-Agent Communication via Identity-Aware 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 IDEAL, an identity-aware heterogeneous message-passing extension for GNN-based multi-agent communication that colors the central agent of its ego network to make agent representations more distinguishable.
Problem
Standard message-passing GNNs used in multi-agent communication are bounded by the 1-WL test, so agents with different neighborhood structures get indistinguishable representations and homogeneous behaviors.
Approach
Extract the L-hop ego network around each agent, color the central agent, and apply separate message-encoding functions to the center and other agents (heterogeneous message passing), which is plug-in compatible with existing GNN-based communication protocols without changing their architecture.
Key results
- On Predator-Prey 20x20, MAGIC-IDEAL needs 28.45 steps versus 33.12 for MAGIC, and TarMAC-IDEAL 31.57 versus 36.22 for TarMAC.
- On Traffic Junction medium, IC3Net-IDEAL reaches 92.15 success versus 87.83 for IC3Net; CommNet-IDEAL reaches 65.27 versus 53.62 for CommNet.
- On SMAC MMM2, DICG-IDEAL achieves 89.23 win rate versus 83.45 for DICG; on Battle, LSC-IDEAL gets 252 kills versus 247 for LSC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add IDEAL to TarMAC and MAGIC | TarMAC and MAGIC · Predator-Prey 10 × 10 | average steps: TarMAC: 13.26 ± 0.10 -> 10.94 ± 0.07; MAGIC: 12.81 ± 0.05 -> 9.72 ± 0.03 | Helped |
| Add IDEAL to TarMAC and MAGIC | TarMAC and MAGIC · Predator-Prey 20 × 20 | average steps: TarMAC: 36.22 ± 0.95 -> 31.57 ± 0.62; MAGIC: 33.12 ± 0.17 -> 28.45 ± 0.11 | Helped |
| Add IDEAL to CommNet and IC3Net | CommNet and IC3Net · Traffic Junction Medium | success rate: CommNet: 53.62 ± 13.81 -> 65.27 ± 11.07; IC3Net: 87.83 ± 3.06 -> 92.15 ± 2.63 | Helped |
| Add IDEAL to CommNet and IC3Net | CommNet and IC3Net · Traffic Junction Hard | success rate: CommNet: 51.56 ± 2.37 -> 67.04 ± 4.28; IC3Net: 73.26 ± 8.72 -> 82.45 ± 6.81 | Helped |
| Add IDEAL to DGN and LSC | DGN and LSC · Battle | kills: DGN: 216 ± 6 -> 225 ± 8; LSC: 247 ± 5 -> 252 ± 5 | Helped |
| Add IDEAL to DGN and LSC | DGN and LSC · Battle | K/D ratio: DGN: 2.32 ± 0.16 -> 2.41 ± 0.19; LSC: 2.60 ± 0.21 -> 2.68 ± 0.23 | Helped |
| Add IDEAL to DGN and LSC | DGN and LSC · Battle | mean reward: DGN: 0.92 ± 0.17 -> 0.96 ± 0.21; LSC: 1.12 ± 0.03 -> 1.16 ± 0.02 | Helped |
| Add IDEAL to G2ANet and DICG | G2ANet and DICG · SMAC MMM2 | win rate: G2ANet: 80.24 ± 4.37 -> 85.31 ± 3.62; DICG: 83.45 ± 5.42 -> 89.23 ± 3.32 | Helped |
| Add IDEAL to G2ANet and DICG | G2ANet and DICG · SMAC 1c3s5z | win rate: G2ANet: 91.25 ± 1.73 -> 94.12 ± 0.94; DICG: 93.46 ± 1.26 -> 95.13 ± 1.13 | Helped |
Limitations
- Improvements on some benchmarks are modest and error bars overlap with baselines (e.g., Battle, Traffic Junction medium for IC3Net).
- The link between communication expressiveness and downstream real-world performance is left for future work.
Experimental setup
Datasets: Predator-Prey Traffic Junction Battle (MAgent) SMAC
Benchmarks: Predator-Prey Traffic Junction Battle SMAC
Metrics: average steps to complete episode success rate win rate kills kill-death ratio mean reward
Baselines: TarMAC MAGIC CommNet IC3Net DGN LSC G2ANet DICG
Builds on: ID-GNN graph decision network MAGIC TarMAC CommNet
Closest papers by meaning
- Learning Multi-Agent Communication from Graph Modeling Perspective (ICLR 2024 · similarity 0.89)
- Learning Efficient and Robust Multi-Agent Communication via Graph Information Bottleneck (AAAI 2024 · similarity 0.89)
- G-Designer: Architecting Multi-agent Communication Topologies via Graph Neural Networks (ICML 2025 · similarity 0.88)
- Multi-Agent Reinforcement Learning with Communication-Constrained Priors (NeurIPS 2025 · similarity 0.87)
- Bridging Training and Execution via Dynamic Directed Graph-Based Communication in Cooperative Multi-Agent Systems (AAAI 2025 · similarity 0.87)
- Learning Multi-Agent Communication with Contrastive Learning (ICLR 2024 · 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.