Research map / Multi-agent systems and games / Cooperative multi-agent RL
Robust Communicative Multi-Agent Reinforcement Learning with Active Defense
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 ADMAC, an active-defense multi-agent communication framework that estimates per-message reliability and weights each message's action-preference contribution, making communicative MARL robust to noisy and adversarial messages.
Problem
Communicative multi-agent reinforcement learning is vulnerable to noisy or adversarially perturbed messages, and existing defenses are passive (treat all messages equally), which trades off performance against robustness.
Approach
Agents estimate the reliability of each received message from their own observation and hidden state, and a decomposable policy net adds each message's action-preference vector scaled by that reliability weight, so unreliable messages can be down-weighted before the softmax decision.
Key results
- ADMAC achieves the best overall performance across three communication-critical tasks under four attack types, with lower timesteps than TARMAC, AT, and AME under attack.
- AME performs poorly without attacks because it averages out all messages, while ADMAC has better baseline performance than TARMAC in Predator Prey.
- An ideal reliability estimator (DPN+IRE) gives the best robust performance, and the gap to DPN+RE is explained by the estimator's recall and precision.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use active defense with ADMAC instead of the evaluated alternatives | communicative MARL · Food Collector, Predator Prey, Treasure Hunt under attacks | average timesteps | Helped |
| use adversarial training instead of TARMAC | communicative MARL · Food Collector, Predator Prey, Treasure Hunt under strong attacks | average timesteps | Helped |
| use adversarial training instead of TARMAC | communicative MARL · Food Collector, Predator Prey, Treasure Hunt without attacks | average timesteps | Hurt |
| use AME message consensus instead of TARMAC | communicative MARL · Food Collector, Predator Prey, Treasure Hunt under attacks | average timesteps | Mixed |
| replace TARMAC message aggregation with the decomposable policy net (ablation) | ADMAC · Food Collector, Predator Prey, Treasure Hunt under attacks | average timesteps | Helped |
| add a reliability estimator to the decomposable policy net (ablation) | DPN · Food Collector, Predator Prey, Treasure Hunt under attacks, especially adversarial attacks | average timesteps | Helped |
| replace the learned reliability estimator with an ideal estimator achieving 100% classification accuracy (ablation) | DPN+RE · Food Collector, Predator Prey, Treasure Hunt under attacks | average timesteps | Helped |
Surprising findings
- AME, the consensus-based defense, degrades baseline performance severely without any attack.
- The decomposable structure alone (DPN only) sometimes gives a little robustness with no defense training.
Limitations
- Performance drops when messages carry unique information that cannot be checked for reliability against local observations.
- Evaluated only on discrete action spaces; continuous action spaces left to future work.
Experimental setup
Datasets: Food Collector Predator Prey Treasure Hunt
Benchmarks: Food Collector Predator Prey Treasure Hunt
Metrics: average timesteps to task completion recall precision
Baselines: TARMAC Adversarial Training (AT) Ablated Message Ensemble (AME)
Builds on: TARMAC Adversarial training REINFORCE noisy learning
Closest papers by meaning
- Multi-Agent Reinforcement Learning with Communication-Constrained Priors (NeurIPS 2025 · similarity 0.89)
- Learning Efficient and Robust Multi-Agent Communication via Graph Information Bottleneck (AAAI 2024 · similarity 0.87)
- Byzantine Robust Cooperative Multi-Agent Reinforcement Learning as a Bayesian Game (ICLR 2024 · similarity 0.87)
- Empirical Study on Robustness and Resilience in Cooperative Multi-Agent Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
- Exploring the Efficacy of Multi-Agent Reinforcement Learning for Autonomous Cyber Defence: A CAGE Challenge 4 Perspective (AAAI 2025 · similarity 0.86)
- Exponential Topology-enabled Scalable Communication in Multi-agent Reinforcement Learning (ICLR 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.