Research map / Reinforcement learning / Visual and representation RL
Honor of Kings Arena: an Environment for Generalization in Competitive Reinforcement Learning
Hua Wei, Jingxiao Chen, Xiyang Ji, Hongyang Qin, Minwen Deng, Siqin Li, Liang Wang, Weinan Zhang, Yong Yu, Liu Linc, Lanxiao Huang, Deheng Ye and 2 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 |
|---|---|---|---|
| Increase the number of CPU cores for PPO self-play data collection | PPO self-play · Diaochan vs. BT | training hours to beat BT: 6.16$\pm$0.009 -> 0.89$\pm$0.06 | Improved |
| Transfer a policy trained on Diaochan vs. Diaochan to other opponent heroes | PPO · Diaochan vs. different BT opponent heroes | win rate | Hurt |
| Train on five opponent-hero tasks instead of only Diaochan vs. Diaochan | PPO · Diaochan vs. 20 BT opponent-hero tasks | win rate | Improved |
| Distill a policy from models trained on five opponent-hero tasks | PPO · Diaochan vs. 20 BT opponent-hero tasks | win rate | Improved |
| Transfer a policy trained on Diaochan vs. Diaochan to control other target heroes | PPO · Different target heroes vs. Diaochan BT | win rate | Hurt |
| Train on five target-hero tasks instead of only Diaochan vs. Diaochan | PPO · 20 target-hero tasks vs. Diaochan BT | win rate | Improved |
| Distill a policy from models trained on five target-hero tasks | PPO · 20 target-hero tasks vs. Diaochan BT | win rate | Improved |
| Add dual clipping to PPO (ablation) | PPO · Diaochan vs. Diaochan | performance | Improved |
| Remove legal action masking (ablation) | PPO · Diaochan vs. Diaochan | performance | Hurt |
| Add an LSTM to the policy network (ablation) | PPO · Diaochan vs. Diaochan | performance | Improved |
| Train Model B against a frozen Model A rather than BT | Model B · Model B vs. BT | win rate: 0.91$\pm$0.03 vs. 0.81$\pm$0.25 | Hurt |
| Use Elo score instead of a single-opponent win rate to evaluate agents | Honor of Kings Arena agent evaluation · Model A, Model B, and BT | Elo score: Model A 2372; Model B 1186; BT 0 | Improved |
Closest papers by meaning
- FightLadder: A Benchmark for Competitive Multi-Agent Reinforcement Learning (ICML 2024 · similarity 0.88)
- On Generalization Across Environments In Multi-Objective Reinforcement Learning (ICLR 2025 · similarity 0.87)
- The Generalization Gap in Offline Reinforcement Learning (ICLR 2024 · similarity 0.86)
- An Open-Ended Learning Framework for Opponent Modeling (AAAI 2025 · similarity 0.85)
- LOQA: Learning with Opponent Q-Learning Awareness (ICLR 2024 · similarity 0.85)
- Learning Diverse Risk Preferences in Population-Based Self-Play (AAAI 2024 · 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.
