Research map / Reinforcement learning / Visual and representation RL
Discovering Hierarchical Achievements in Reinforcement Learning via Contrastive Learning
Seungyong Moon, Junyoung Yeom, Bumsoo Park, Hyun Oh Song
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 |
|---|---|---|---|
| Add achievement distillation (contrastive auxiliary learning) | PPO · Crafter 1M steps | score (%): 15.60 -> 21.79 | Improved |
| Add intra-trajectory achievement prediction (ablation) | PPO · Crafter 1M steps | score (%): 15.60 -> 19.02 | Improved |
| Add cross-trajectory achievement matching (ablation) | PPO + intra-trajectory prediction · Crafter 1M steps | score (%): 19.02 -> 20.36 | Improved |
| Add memory (ablation) | PPO + prediction + matching · Crafter 1M steps | score (%): 20.36 -> 21.79 | Improved |
| Shrink model to 1M parameters (channels [16, 32, 32], hidden 256) | Ours · Crafter 1M steps | score (%): 21.79 -> 17.07 | Hurt |
| Add achievement distillation | PPO · Crafter 1M steps | classification accuracy (%): 73.6 | Improved |
| Add achievement distillation | QR-DQN · Crafter 1M steps | score (%): 4.14 -> 8.07 | Improved |
| Add achievement distillation | PPO · Procgen Heist | score: 29.6 -> 71.0 | Improved |
| Add achievement distillation | PPO · MiniGrid | score: 3.33 -> 8.04 | Improved |
| Use PPO with recent implementation practices instead of DreamerV3 | PPO · Crafter 1M steps | score (%): 14.77 -> 15.60 | Improved |
| Use PPO instead of LSTM-SPCNN (recurrent object-centric network) | PPO · Crafter 1M steps | score (%): 11.67 -> 15.60 | Improved |
| Our method (9M params) versus MuZero + SPR with pre-training | achievement distillation · Crafter 1M steps | score (%): 16.4 -> 21.79 | Improved |
Closest papers by meaning
- Enhancing Exploration and Exploitation in Hierarchical Reinforcement Learning with Subgoal Graph Learning (AAAI 2026 · similarity 0.86)
- Sub-policy Adaptation for Hierarchical Reinforcement Learning (ICLR 2020 · similarity 0.86)
- ContraDiff: Planning Towards High Return States via Contrastive Learning (ICLR 2025 · similarity 0.86)
- Emergent Hierarchical Reasoning in LLMs through Reinforcement Learning (ICLR 2026 · similarity 0.86)
- DISCOVER: Automated Curricula for Sparse-Reward Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- Discovering Temporally-Aware Reinforcement Learning Algorithms (ICLR 2024 · similarity 0.85)
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.
