Research map / Multi-agent systems and games / Cooperative multi-agent RL
PMAC: Personalized Multi-Agent Communication
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 PMAC, a multi-agent RL communication method that learns peer-to-peer directed communication topologies and personalized message sending/receiving using only MLPs with linear complexity.
Problem
Multi-agent reinforcement learning communication struggles to send and receive personalized messages and to build peer-to-peer topologies, since existing graph/attention methods incur O(n^2) cost in the number of agents.
Approach
Models the directed communication graph via a low-rank factorization of the adjacency matrix (sum of products of sender and receiver MLP embeddings, sampled with Gumbel-softmax), generates receiver-specific messages with a single forward pass using an identity-ID trick (Theorem 1), and applies affine receiving that can be swapped with aggregation, giving linear complexity in agents.
Key results
- On hard Traffic Junction, PMAC reaches 85.1% success rate vs 79.8% for MAGIC, 79.1% for DGN, 78.6% for TarMAC and 60.6% for CommNet.
- On Google Research Football 3v1 with Keeper, PMAC achieves the best final success rate among five methods, while CommNet reaches only about 40% win rate.
- Computational time per 1000 episodes on TJ with 20 agents is lower for PMAC than MAGIC, DGN, TarMAC and CommNet, with the low-rank topology module contributing most to speedup.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PMAC instead of MAGIC | multi-agent communication · hard-TJ | success rate: 79.8 ± 7.7 -> 85.1 ± 6.6 | Helped |
| Use PMAC instead of DGN | multi-agent communication · hard-TJ | success rate: 79.1 ± 9.8 -> 85.1 ± 6.6 | Helped |
| Use PMAC instead of TarMAC | multi-agent communication · hard-TJ | success rate: 78.6 ± 2.9 -> 85.1 ± 6.6 | Helped |
| Use PMAC instead of CommNet | multi-agent communication · hard-TJ | success rate: 60.6 ± 10.5 -> 85.1 ± 6.6 | Helped |
| Use PMAC instead of MAGIC | multi-agent communication · Cooperative Navigation | reward | Helped |
| Use PMAC instead of DGN and MAGIC | multi-agent communication · Predator-Prey | mean reward | Helped |
| Use PMAC instead of the other four methods | multi-agent communication · Google Research Football 3 versus 1 with Keeper | success rate | Helped |
| Set communication regularization coefficient η to 0.001 rather than the other tested coefficients (ablation) | PMAC · Predator-Prey | reward | Helped |
| Use a fully connected topology instead of personalized communication topology (ablation) | PMAC · Predator-Prey | reward | Hurt |
| Send the same messages to different receivers instead of personalized messages (ablation) | PMAC · Predator-Prey | reward | Hurt |
| Aggregate received messages without considering the sender (ablation) | PMAC · Predator-Prey | reward | Hurt |
| Remove low-rank approximation for communication topology (ablation) | PMAC · Predator-Prey | computational efficiency | Hurt |
Surprising findings
- PMAC without personalized message sending converges faster with lower variance, but reaches a lower final reward than full PMAC on predator-prey.
- Even with global state access in GRF, communication via PMAC improves performance.
Limitations
- Personalization relies on agent ID one-hot encodings, a simple form of identification.
- Results reported over a small number of runs (3 or 5 seeds) with sizeable variance on some tasks.
- PMAC initially underperforms DGN and MAGIC on predator-prey before overtaking them.
Experimental setup
Datasets: Traffic Junction (hard) Cooperative Navigation (MPE) Predator-Prey (MPE) Google Research Football 3 vs 1 with Keeper
Benchmarks: Traffic Junction (hard) Cooperative Navigation Predator-Prey Google Research Football
Metrics: success rate mean reward win rate wall-clock time per 1000 episodes
Baselines: CommNet TarMAC DGN MAGIC
Builds on: DIAL TarMAC MAGIC DGN IC3Net
Closest papers by meaning
- Exponential Topology-enabled Scalable Communication in Multi-agent Reinforcement Learning (ICLR 2025 · similarity 0.86)
- Efficient Communication in Multi-Agent Reinforcement Learning with Implicit Consensus Generation (AAAI 2025 · similarity 0.86)
- T2MAC: Targeted and Trusted Multi-Agent Communication through Selective Engagement and Evidence-Driven Integration (AAAI 2024 · similarity 0.86)
- Expressive Multi-Agent Communication via Identity-Aware Learning (AAAI 2024 · similarity 0.86)
- Multi-Agent Reinforcement Learning with Communication-Constrained Priors (NeurIPS 2025 · similarity 0.86)
- Learning Multi-Agent Communication from Graph Modeling Perspective (ICLR 2024 · 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.