Research map / Reinforcement learning / Offline RL
Hokoff: Real Game Dataset from Honor of Kings and its Offline Reinforcement Learning Benchmarks
Yun Qu, Boyuan Wang, Jianzhun Shao, Yuhang Jiang, Chen Chen, Zhenbin Ye, Lin Liu, Yang Jun Feng, Lin Lai, Hongyang Qin, Minwen Deng, Juchao Zhuo and 6 more
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 |
|---|---|---|---|
| use QMIX+CQL instead of CQL (MARL structure for structured action space) | CQL · HoK1v1 hard_expert | winning rate: 0.30±0.01 -> 0.44±0.05 | Improved |
| use QMIX+CQL instead of CQL | CQL · HoK1v1 norm_general | winning rate: 0.20±0.04 -> 0.32±0.03 | Improved |
| use TD3-based TD3+BC instead of BC | BC · HoK1v1 norm_expert | winning rate: 0.64±0.01 -> 0.03±0.01 | Hurt |
| use OMAR instead of IND+CQL | IND+CQL · HoK3v3 norm_expert | winning rate: 0.84±0.06 -> 0.39±0.16 | Hurt |
| add communication (COMM+CQL vs IND+CQL) (ablation) | IND+CQL · HoK3v3 norm_mixed | winning rate: 0.46±0.12 -> 0.48±0.12 | Improved |
| add communication (COMM+CQL vs IND+CQL) (ablation) | IND+CQL · HoK3v3 norm_multi_hero | winning rate: 0.3±0.07 -> 0.31±0.2 | No clear effect |
| use CTDE (MAICQ) instead of independent learning (IND+ICQ) (ablation) | IND+ICQ · HoK3v3 norm_expert | winning rate: 0.65±0.12 -> 0.61±0.09 | Hurt |
| use CTDE (MAICQ) instead of independent learning (IND+ICQ) (ablation) | IND+ICQ · HoK3v3 hard_general | winning rate: 0.31±0.09 -> 0.14±0.06 | Hurt |
| hard-level instead of norm-level dataset | QMIX+CQL · HoK1v1 expert | winning rate: 0.67±0.01 -> 0.44±0.05 | Hurt |
| hard-level instead of norm-level dataset | IND+CQL · HoK3v3 expert | winning rate: 0.84±0.06 -> 0.67±0.04 | Hurt |
| expert partner instead of stupid partner | IND+CQL · HoK3v3 heterogeneous | winning rate: 0.24±0.17 -> 0.57±0.04 | Improved |
| multi-task training instead of generalization datasets | QMIX+CQL · HoK1v1 norm | winning rate: 0.32±0.03 -> 0.41±0.02 | Improved |
Closest papers by meaning
- Scaling Offline Model-Based RL via Jointly-Optimized World-Action Model Pretraining (ICLR 2025 · similarity 0.89)
- Offline Fictitious Self-Play for Competitive Games (AAAI 2026 · similarity 0.88)
- Dispelling the Mirage of Progress in Offline MARL through Standardised Baselines and Evaluation (NeurIPS 2024 · similarity 0.87)
- The Generalization Gap in Offline Reinforcement Learning (ICLR 2024 · similarity 0.87)
- FightLadder: A Benchmark for Competitive Multi-Agent Reinforcement Learning (ICML 2024 · similarity 0.87)
- A Clean Slate for Offline Reinforcement Learning (NeurIPS 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.
