Research map / Reinforcement learning / Offline RL
Offline Reinforcement Learning with Adaptive Feature Fusion
Tieru Wang, Kunbao Wu, Guoshun Nan
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 |
|---|---|---|---|
| set context length to 1 (single-step DT) in toy task (ablation) | DT · toy trajectory-stitching task | success rate (%): 0.0 -> 100.0 | Improved |
| add Q-value guidance with eta = 0.01 to DT-style model (K=3) (ablation) | QT · toy task | success rate (%): 0.1 | No clear effect |
| increase QT eta to 10.0 in toy task (ablation) | QT · toy task | success rate (%): 0.1 -> 96.2 | Improved |
| learn adaptive fusion coefficient instead of fixed alpha-hat = 0 (ablation) | QDFFDT · Maze2D+AntMaze | average score: 86.9 -> 105.3 | Improved |
| learn adaptive fusion instead of fixed alpha-hat = 1 (sequence only) (ablation) | QDFFDT · Maze2D+AntMaze | average score: 101.2 -> 105.3 | Improved |
| learn fusion instead of fixed alpha-hat = 0.25 (ablation) | QDFFDT · Maze2D+AntMaze | average score: 80.4 -> 105.3 | Improved |
| add Q-augmented dual-feature fusion to QT | QDFFDT vs QT · D4RL Gym average | normalized score: 89.8 -> 94.3 | Improved |
| add adaptive fusion over QT on AntMaze | QDFFDT vs QT · D4RL AntMaze average | normalized score: 38.9 -> 88.3 | Improved |
| use concatenation fusion instead of addition (hopper-medium-replay) (ablation) | QDFFDT · hopper-medium-replay-v2 | normalized score: 102.7 -> 103.9 | No clear effect |
| use concatenation fusion instead of addition (walker2d-medium-replay) (ablation) | QDFFDT · walker2d-medium-replay-v2 | normalized score: 101.0 -> 100.3 | No clear effect |
| use n-step Bellman instead of 1-step with K=20 (ablation) | QDFFDT · halfcheetah-medium-v2 | normalized score: 65.7 -> 55.6 | Hurt |
| use n-step Bellman instead of 1-step with K=20 (ablation) | QDFFDT · halfcheetah-medium-replay-v2 | normalized score: 62.0 -> 38.4 | Hurt |
Closest papers by meaning
- Adaptive $Q$-Aid for Conditional Supervised Learning in Offline Reinforcement Learning (NeurIPS 2024 · similarity 0.89)
- Critic-Guided Decision Transformer for Offline Reinforcement Learning (AAAI 2024 · similarity 0.89)
- When should we prefer Decision Transformers for Offline Reinforcement Learning? (ICLR 2024 · similarity 0.89)
- Q-value Regularized Transformer for Offline Reinforcement Learning (ICML 2024 · similarity 0.88)
- Adaptive Neighborhood-Constrained Q Learning for Offline Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- In-Dataset Trajectory Return Regularization for Offline Preference-based Reinforcement Learning (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.
