Research map / Reinforcement learning / Offline RL
Offline Reinforcement Learning with Universal Horizon Models
Hojun Chung, Junseo Lee, Songhwai Oh
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 lambda scheduling for UHM training (ablation) | UHM · OGBench (antmaze-giant-navigate-1, puzzle-3x3-play-2, cube-single-play-1) | success rate | Improved |
| add terminal state handling (absorbing terminal states, no bootstrapping) (ablation) | UHM · OGBench (Hmaze-medium-navigate-1, puzzle-3x3-play-4, cube-single-play-2) | success rate | Improved |
| winsorize horizon at quantile q=0.1 or 0.2 instead of q=1e-8 (ablation) | UHM · Long-horizon OGBench tasks | success rate | Improved |
| behavior mixing coefficient beta 0.3 versus 0.0 and 1.0 (ablation) | UHM · Representative standard OGBench tasks | average success rate: 0.63 / 0.66 / 0.59 for beta 0.0 / 0.3 / 1.0 | Mixed |
| use UHM n-step future prediction instead of GHM geometric horizon model (ablation) | Model-based value expansion · OGBench 25 long-horizon tasks | average success rate (%): 16 -> 22 | Improved |
| use UHM instead of GHM on standard tasks (ablation) | Model-based value expansion · OGBench 50 standard tasks | average success rate (%): 48 -> 55 | Improved |
| use UHM on-policy imagined rollouts instead of single-step dynamics model rollouts MBTD(λ) (ablation) | TD(λ) value learning · OGBench 50 standard tasks | average success rate (%): 45 -> 55 | Improved |
| use dataset trajectories (DTD(λ)) instead of imagined on-policy trajectories (ablation) | TD(λ) value learning · OGBench 25 noisy tasks | average success rate (%): 39 -> 23 | Hurt |
| use UHM instead of dataset trajectories (DTD(λ)) on long-horizon tasks (ablation) | TD(λ) value learning · OGBench 25 long-horizon tasks | average success rate (%): 13 -> 22 | Improved |
| use n-step return TD instead of one-step TD (ablation) | Offline RL critic · OGBench 50 standard tasks | average success rate (%): ReBRAC† 35 -> DTD(λ) 52 | Improved |
| use single-step dynamics model rollouts for long-horizon imagination | MBTD(λ) · OGBench 25 long-horizon tasks | average success rate (%): 5 vs ReBRAC† 5 | No clear effect |
| model-based methods with uncertainty-based reward penalties | MOPO and MOBILE · OGBench sparse-reward standard tasks | success rate: MOPO 4, MOBILE 10 average | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MOPO — MOPO: Model-based Offline Policy Optimization (NeurIPS 2020) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Horizon Adaptive Offline Policy Learning via Value Stitching (NeurIPS 2026 · similarity 0.90)
- Horizon Reduction Makes RL Scalable (NeurIPS 2025 · similarity 0.90)
- The Edge-of-Reach Problem in Offline Model-Based Reinforcement Learning (NeurIPS 2024 · similarity 0.89)
- Option-aware Temporally Abstracted Value for Offline Goal-Conditioned Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- Scaling Offline Model-Based RL via Jointly-Optimized World-Action Model Pretraining (ICLR 2025 · similarity 0.87)
- Forecasting in Offline Reinforcement Learning for Non-stationary Environments (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.
