Research map / Multi-agent systems and games / Cooperative multi-agent RL
Learning Efficient and Robust Multi-Agent Communication via Graph Information Bottleneck
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 MAGI, a GNN-based multi-agent communication method that applies a graph information bottleneck objective (maximize MI with actions, minimize MI with graph/agent features) on top of QMIX, improving robustness and efficiency under adversarial attacks and noise in SMAC and MAgent.
Problem
Communication learning in GNN-based cooperative multi-agent RL is vulnerable to adversarial attacks and noise on agent features and graph structure, and existing methods use message representations that carry irrelevant information.
Approach
Extends the graph information bottleneck principle to GNN-based MACRL: message representations are trained to maximize mutual information with selected actions while minimizing mutual information with the graph-structured agent input, using layer-wise local-dependence variational bounds (categorical neighbor sampling for structure, Gaussian mixture prior for features) integrated with QMIX.
Key results
- MAGI+A keeps win rates close to clean MAGI under adversarial attacks and noise, while TarMAC+A and MAGIC+A degrade significantly (Figure 3).
- On MMM3 under GN+PGD attack, MAGI reaches 71.34±4.26 win rate versus 56.06±6.32 for MAGIC and 45.42±8.03 for TarMAC (Table 4).
- Removing the IB component (MAGI-IB) drops 2c3s5z win rate to 69.35 from 93.02 for full MAGI (Table 2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add graph information bottleneck optimization (ablation) | MAGI · MMM2 (SMAC) | win rate: 60.07±5.75 -> 82.87±2.93 | Helped |
| add graph information bottleneck optimization (ablation) | MAGI · MMM3 (SMAC) | win rate: 42.77±4.62 -> 57.84±3.96 | Helped |
| add graph information bottleneck optimization (ablation) | MAGI · 8m vs 9m (SMAC) | win rate: 68.12±3.83 -> 92.55±1.78 | Helped |
| add graph information bottleneck optimization (ablation) | MAGI · 1o2r vs 4r (SMAC) | win rate: 71.24±5.49 -> 86.01±3.04 | Helped |
| add graph information bottleneck optimization (ablation) | MAGI · 2c3s5z (SMAC) | win rate: 69.35±4.18 -> 93.02±1.45 | Helped |
| add value decomposition (ablation) | MAGI · MMM2 (SMAC) | win rate: 79.93±3.64 -> 82.87±2.93 | Helped |
| add value decomposition (ablation) | MAGI · MMM3 (SMAC) | win rate: 52.28±3.25 -> 57.84±3.96 | Helped |
| add value decomposition (ablation) | MAGI · 8m vs 9m (SMAC) | win rate: 87.29±2.40 -> 92.55±1.78 | Helped |
| add value decomposition (ablation) | MAGI · 1o2r vs 4r (SMAC) | win rate: 84.64±3.56 -> 86.01±3.04 | Helped |
| add value decomposition (ablation) | MAGI · 2c3s5z (SMAC) | win rate: 89.47±1.90 -> 93.02±1.45 | Helped |
| set the information bottleneck loss weight to λ=0.15 instead of λ=0.10 (ablation) | MAGI · MMM3 (SMAC) | win rate: 57.84±3.96 -> 59.27±3.71 | Helped |
| set the information bottleneck loss weight to λ=0.05 instead of λ=0.10 (ablation) | MAGI · MMM2 (SMAC) | win rate: 82.87±2.93 -> 80.17±3.02 | Hurt |
Limitations
- Evaluated only on SMAC and MAgent simulated environments
- Adversarial attacks are limited to specific attack methods (GN+PGD, IG-JSMA, GUA)
Experimental setup
Datasets: SMAC MAgent
Benchmarks: SMAC MAgent
Metrics: win rate mean reward
Baselines: QMIX TarMAC MAGIC MAGI-IB MAGI-VD
Builds on: QMIX Graph information bottleneck TarMAC MAGIC Graph attention network
Closest papers by meaning
- Learning Multi-Agent Communication from Graph Modeling Perspective (ICLR 2024 · similarity 0.90)
- Bridging Training and Execution via Dynamic Directed Graph-Based Communication in Cooperative Multi-Agent Systems (AAAI 2025 · similarity 0.89)
- Expressive Multi-Agent Communication via Identity-Aware Learning (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 0.88)
- Efficient Communication in Multi-Agent Reinforcement Learning with Implicit Consensus Generation (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.