Research map / Multi-agent systems and games / Cooperative multi-agent RL
TACTIC: Task-Aware Sparse Coordination Graphs for Multi-Task Multi-agent Reinforcement Learning
Kexing Peng, Pengyi Li, tinghuai ma, Jianye HAO
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add trajectory abstraction (VQ-VAE classes) (ablation) | TACTIC · SMACv2 10zerg and 10terran | win rate | Improved |
| add coverage loss (ablation) | TACTIC · SMACv2 10zerg and 10terran | win rate | Improved |
| use dynamic task-conditioned graph instead of random graph (ablation) | TACTIC · SMACv2 | win rate | Improved |
| add trajectory classification and graph modulation together (ablation) | TACTIC · SMACv2 | win rate | Improved |
| use dynamic graph instead of static graph (ablation) | TACTIC · reSurComb3/4 (mirrored tasks) | return | Hurt |
| use dynamic graph instead of static graph (ablation) | TACTIC · SurComb3/4 | return | Improved |
| add full semantic-to-structure coupling instead of only a dynamic graph on TRAMA backbone (ablation) | TACTIC · SMACv2 SurComb3 | reward: 2.38 -> 2.80 | Improved |
| update graph every 5 steps instead of every step at inference (ablation) | TACTIC · SMACv2 SurComb3 | reward: 2.80 -> 2.73 | Mixed |
| update graph every 10 steps instead of every step at inference (ablation) | TACTIC · SMACv2 SurComb3 | reward: 2.80 -> 2.65 | Mixed |
| use static graph per episode instead of per-step adaptive (ablation) | TACTIC · SMACv2 SurComb3 | reward: 2.80 -> 2.50 | Hurt |
| use TACTIC instead of Random policy | TACTIC · SUMO intersection | arrival time (s): 18.60 -> 15.90 | Improved |
| use TACTIC instead of Random policy | TACTIC · SUMO intersection | fuel (mg): 21813.26 -> 19008.64 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- LAGMA — LAGMA: LAtent Goal-guided Multi-Agent Reinforcement Learning (ICML 2024) · better on 4 of 4 shared results
- QMIX — QMIX: Monotonic Value Function Factorisation for Deep Multi-Agent Reinforcement Learning (ICML 2018) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Context-Aware Sparse Deep Coordination Graphs (ICLR 2022 · similarity 0.90)
- Dynamic Deep Factor Graph for Multi-Agent Reinforcement Learning (TPAMI 2025 · similarity 0.88)
- Non-Linear Coordination Graphs (NeurIPS 2022 · similarity 0.88)
- Graph Diffusion for Robust Multi-Agent Coordination (ICML 2025 · similarity 0.88)
- Trajectory-Class-Aware Multi-Agent Reinforcement Learning (ICLR 2025 · similarity 0.87)
- Learning Generalizable Skills from Offline Multi-Task Data for Multi-Agent Cooperation (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.
