Research map / Reinforcement learning / Visual and representation RL
PACE: Parameter Change for Unsupervised Environment Design
Fang YUAN, Junjie Zeng, Qinglun Li, Long Qin, Quanjun Yin, Siqi Shen, Yuxiang Xie, Junqiang Yang
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 |
|---|---|---|---|
| score levels by squared parameter change of a provisional update instead of GAE or MaxMC proxies | PACE (PPO student) · MiniGrid (12 held-out levels) | IQM of solved rate: 0.808 -> 0.964 | Improved |
| score levels by squared parameter change instead of proxy scores | PACE (PPO student) · MiniGrid (12 held-out levels) | Optimality Gap: 0.299 -> 0.172 | Improved |
| score levels by squared parameter change instead of proxy scores | PACE (PPO student) · MiniGrid PerfectMaze (51x51) | solve rate (%): 72.3% vs baselines 26.33-54.17% | Improved |
| score levels by squared parameter change instead of proxy scores | PACE (PPO student) · Craftax-1B | IQM of reward: 0.618 -> 0.694 | Improved |
| score levels by squared parameter change instead of proxy scores | PACE (PPO student) · Craftax-1B | evaluation reward: 31.198 ± 2.439 | Improved |
| use PACE scoring instead of PLR/ACCEL on easy MiniGrid levels (FourRooms, Crossing) | PACE (PPO student) · MiniGrid FourRooms and Crossing | solved rate: all methods above 84% on FourRooms | No clear effect |
| use PACE scoring instead of PLR⊥ on 16RoomsFewerDoors | PACE (PPO student) · MiniGrid 16RoomsFewerDoors | solved rate: 0.972 -> 0.728 | Hurt |
| use PACE scoring instead of PLR on large PerfectMaze(M) level | PACE (PPO student) · MiniGrid PerfectMaze(M) | solved rate: 0.373 -> 0.359 | No clear effect |
| use parameter-change score instead of PLR/PLR⊥/ACCEL scores | PACE · MiniGrid score stability | Mean CV (lower better): 0.0055 (PLR⊥) -> 0.0010 | Improved |
| use parameter-change score instead of PLR/PLR⊥/ACCEL scores | PACE · Craftax score stability | Mean CV (lower better): 0.6105 (PLR) -> 0.5412 | Improved |
| regret-based curation (ACCEL) versus domain randomization in Craftax | ACCEL · Craftax-1B | evaluation reward | No clear effect |
Closest papers by meaning
- Improving Regret Approximation for Unsupervised Dynamic Environment Generation (NeurIPS 2025 · similarity 0.92)
- No Regrets: Investigating and Improving Regret Approximations for Curriculum Discovery (NeurIPS 2024 · similarity 0.89)
- Adversarial Environment Design via Regret-Guided Diffusion Models (NeurIPS 2024 · similarity 0.89)
- Improving Environment Novelty Quantification for Effective Unsupervised Environment Design (NeurIPS 2024 · similarity 0.88)
- Unsupervised Training Sequence Design: Efficient and Generalizable Agent Training (AAAI 2024 · similarity 0.88)
- Marginal Benefit Driven RL Teacher for Unsupervised Environment Design (AAAI 2025 · similarity 0.87)
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.
