Research map / Reinforcement learning / Visual and representation RL
Hierarchical Skills for Efficient Exploration
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 three-level hierarchical goal-space selection over a single full goal space (ablation) | HSD-3 · Walker Stairs | average return: 15.2 -> 17.2 | Helped |
| add three-level hierarchical goal-space selection over a single full goal space (ablation) | HSD-3 · Walker GoalWall | average return: 0.4 -> 0.9 | Helped |
| add three-level hierarchical goal-space selection over a single full goal space (ablation) | HSD-3 · Walker Limbo | average return: 11.0 -> 11.7 | No clear effect |
| add three-level hierarchical goal-space selection over a single full goal space (ablation) | HSD-3 · Walker Hurdles | average return: 15.0 -> 15.3 | No clear effect |
| replace learned high-level goal-space policy with a per-task bandit (ablation) | HSD-3 · Walker Hurdles | average return: 15.3 -> 2.6 | Hurt |
| replace learned high-level goal-space policy with a per-task bandit (ablation) | HSD-3 · Walker Stairs | average return: 17.2 -> 0.5 | Hurt |
| use skill pre-training instead of no pre-training (SAC) | HSD-3 · Walker Hurdles | average return: 2.2 -> 15.3 | Helped |
| use skill pre-training instead of no pre-training (SAC) | HSD-3 · PoleBalance | average return: 866.8 -> 876.0 | No clear effect |
| use HSD pre-training instead of DIAYN-C pre-training | HSD-3 · Walker Stairs | average return: 12.5 -> 17.2 | Helped |
| use HSD-3 instead of best per-task fixed goal space chosen in hindsight | HSD-3 · Walker Limbo | average return: 12.2 -> 11.7 | No clear effect |
| use best fixed goal space per task rather than the most general one | fixed Walker skills · Walker benchmark tasks | average return | Mixed |
| control all features (most general skill) instead of task-specific skill | fixed Walker skills · Walker benchmark tasks | average return | Hurt |
Closest papers by meaning
- Diversity is All You Need: Learning Skills without a Reward Function (ICLR 2019 · similarity 0.90)
- Dynamics-Aware Unsupervised Discovery of Skills (ICLR 2020 · similarity 0.89)
- Training Transition Policies via Distribution Matching for Complex Tasks (ICLR 2022 · similarity 0.89)
- Exploration by Learning Diverse Skills through Successor State Representations (NeurIPS 2024 · similarity 0.88)
- When Do Skills Help Reinforcement Learning? A Theoretical Analysis of Temporal Abstractions (ICML 2024 · similarity 0.88)
- Periodic Skill Discovery (NeurIPS 2025 · similarity 0.88)
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.