Research map / Reinforcement learning / Visual and representation RL
Beyond Fixed Tasks: Unsupervised Environment Design for Task-Level Pairs
Daniel Furelos-Blanco, Charles Pert, Frederik Kelbel, Alex F. Spies, Alessandra Russo, Michael Dennis
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 PLR-perp or ACCEL curricula instead of domain randomization | task-level pair training · Minigrid task-level pairs, independent sampling | CVaR / test solve rate: 2.7% solvable problems per batch; DR solves almost no problems | Improved |
| add level-conditioned problem sampling (83.4% solvable) for DR (ablation) | DR · Hand-designed test set | CVaR and test performance: 2.7% -> 83.4% solvable; DR matches ≈91.6% | Improved |
| add level-conditioned problem sampling for PLR-perp and ACCEL (ablation) | PLR-perp, ACCEL · Hand-designed test set | CVaR and test performance: ≈1--3% gains | No clear effect |
| use random walk-based DAG task sampler instead of sequential (ablation) | PLR-perp · Hand-designed test set | solve rate: drop by ≈35% compared to the sequential setting | Hurt |
| add both level and task edits (vs disabling one type) (ablation) | ACCEL · Minigrid task-level pairs | performance | Improved |
| use single-edit sequences instead of 7-10 edits (ablation) | ACCEL · Minigrid task-level pairs | performance: ≈40% drop | Hurt |
| use 3-edit sequences instead of 7-10 edits (ablation) | ACCEL · Minigrid task-level pairs | performance: ≈15% drop | Hurt |
| use 20-edit sequences instead of 7-10 edits (ablation) | ACCEL · Minigrid task-level pairs | performance: ≈2% variation | No clear effect |
| add mutations (ACCEL) on top of curation (PLR-perp) | PLR-perp · Hand-designed test set | convergence speed | Improved |
Closest papers by meaning
- CLUTR: Curriculum Learning via Unsupervised Task Representation Learning (ICML 2023 · similarity 0.90)
- Replay-Guided Adversarial Environment Design (NeurIPS 2021 · similarity 0.89)
- MAESTRO: Open-Ended Environment Design for Multi-Agent Reinforcement Learning (ICLR 2023 · similarity 0.89)
- Unsupervised Training Sequence Design: Efficient and Generalizable Agent Training (AAAI 2024 · similarity 0.88)
- No Regrets: Investigating and Improving Regret Approximations for Curriculum Discovery (NeurIPS 2024 · similarity 0.87)
- Improving Regret Approximation for Unsupervised Dynamic Environment Generation (NeurIPS 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.
