Research map / Multi-agent systems and games / Human-AI and goal-conditioned
From Few to More: Large-Scale Dynamic Multiagent Curriculum Learning
Weixun Wang, Tianpei Yang, Yong Liu, Jianye Hao, Xiaotian Hao, Yujing Hu, Yingfeng Chen, Changjie Fan, Yang Gao
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 DyMA-CL with Max aggregation | IQL · MAgent 50 vs 50 | Survivors: 0.54$\pm$2.65 -> 20.35$\pm$4.87 | Improved |
| Add DyMA-CL with Max aggregation | IQL · MAgent 50 vs 50 | Kill count: 31.83$\pm$8.52 -> 50$\pm$0 | Improved |
| Add DyMA-CL with Sum aggregation | PPO · MAgent 50 vs 50 | Kill count: 1.06$\pm$1.12 -> 38.25$\pm$7.33 | Improved |
| Add DyMA-CL with Max aggregation | A2C · MAgent 50 vs 50 | Survivors: 3.65$\pm$5.58 -> 16.5$\pm$10.92 | Improved |
| Add DyMA-CL with Max aggregation | A2C · MAgent 50 vs 50 | Kill count: 36.4$\pm$13.79 -> 47.71$\pm$5.61 | Improved |
| Add DyMA-CL with Mean aggregation | ACER · MAgent 50 vs 50 | Kill count: 4.84$\pm$2.8 -> 12.62$\pm$3.68 | Improved |
| Use DyMA-CL with model reload instead of learning from scratch | IQL and VDN · StarCraft II 10I and 15I | average win rate | Improved |
| Train from scratch rather than use DyMA-CL | IQL · StarCraft II 15I | average win rate: 7.5e+6 steps; 0 | Hurt |
| Use SUM aggregation rather than MAX or MEAN (ablation) | DyAN with IQL and VDN · StarCraft II 5I, 10I, and 15I | average win rate | Improved |
| Replace the GNN aggregation operator with MAX (ablation) | DyAN · MAgent 50 vs 50 | Survivors: IQL: 2.33$\pm$2.9 (Mean) vs 20.35$\pm$4.87 (Max); A2C: 6.8$\pm$7.39 (Mean) vs 16.5$\pm$10.92 (Max) | Mixed |
| Use curriculum distillation instead of learning from scratch | IQL and VDN · StarCraft II 10I and 15I | average win rate | Improved |
| Reuse replay buffers from earlier curricula | IQL and VDN · StarCraft II 10I and 15I | average win rate | Improved |
Closest papers by meaning
- PORTAL: Automatic Curricula Generation for Multiagent Reinforcement Learning (AAAI 2024 · similarity 0.89)
- Learning to Teach in Cooperative Multiagent Reinforcement Learning (AAAI 2019 · similarity 0.87)
- Curriculum Design for Trajectory-Constrained Agent: Compressing Chain-of-Thought Tokens in LLMs (NeurIPS 2025 · similarity 0.86)
- Learning Progress Driven Multi-Agent Curriculum (ICML 2025 · similarity 0.86)
- Dr. MAS: Stable Reinforcement Learning for Multi-Agent LLM Systems (NeurIPS 2026 · similarity 0.85)
- WebRL: Training LLM Web Agents via Self-Evolving Online Curriculum Reinforcement Learning (ICLR 2025 · similarity 0.84)
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.
