Research map / Reinforcement learning / Offline RL
State Proficiency-Based Adaptive Fine-Tuning for Offline-to-Online Reinforcement Learning
Songlin Li, Wei Xiao, Hao Wu, Xiaodan Zhang, Daolong An, Shuai Lü
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 |
|---|---|---|---|
| classify samples by action comparison instead of random classification with matched ratio (ablation) | SPA · D4RL Locomotion (8 medium-level tasks) | total normalized score: 910.1 -> 961.6 | Improved |
| use dual thresholds (offline policy for offline data, historical optimal policy for online data) instead of a single high threshold (ablation) | SPA · walker2d-medium | training stability, constraint ratio on offline data: 43.01 -> 52.07 | Improved |
| use dual thresholds instead of a single low threshold (ablation) | SPA · halfcheetah-medium-replay | final performance, constraint ratio on online data: online constraint ratio 59.27 -> 33.79 | Improved |
| add value-function-based classification correction (ablation) | SPA · walker2d-medium-replay | precision of high proficiency classification (%): 51.70 -> 60.22 | Improved |
| add value-based correction rather than randomly reclassifying the same proportion of samples (ablation) | SPA · walker2d-medium-replay | precision (%): 48.15 -> 60.22 | Improved |
| relax policy constraint ratio alone (without accurate correction) (ablation) | SPA · D4RL Locomotion | performance | No clear effect |
| apply SPA to TD3+BC fine-tuning versus baselines | TD3+BC · D4RL Locomotion (12 tasks) | total normalized score: 694.4 -> 1142.7 | Improved |
| apply SPA to TD3+BC fine-tuning on Maze2D | TD3+BC · D4RL Maze2D (6 tasks) | total normalized score: 362.4 -> 1115.8 | Improved |
| fine-tune with SPA on a task whose offline Q-values exploded | SPA · walker2d-random-v2 | normalized score: 4.2 -> 0.2 | Hurt |
| use action-comparison proficiency on halfcheetah-medium compared with random classification (ablation) | SPA · halfcheetah-medium-v2 | normalized score: 60.3 -> 91.7 | Improved |
| use action-comparison proficiency on walker2d-medium compared with random classification (ablation) | SPA · walker2d-medium-v2 | normalized score: 116.3 -> 116.1 | No clear effect |
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.
- Cal-QL — Cal-QL: Calibrated Offline RL Pre-Training for Efficient Online Fine-Tuning (NeurIPS 2023) · 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
- SUF: Stabilized Unconstrained Fine-Tuning for Offline-to-Online Reinforcement Learning (AAAI 2024 · similarity 0.90)
- Efficient Online Reinforcement Learning Fine-Tuning Need Not Retain Offline Data (ICLR 2025 · similarity 0.89)
- Optimistic Critic Reconstruction and Constrained Fine-Tuning for General Offline-to-Online RL (NeurIPS 2024 · similarity 0.88)
- Adaptive Advantage-Guided Policy Regularization for Offline Reinforcement Learning (ICML 2024 · similarity 0.88)
- A Perspective of Q-value Estimation on Offline-to-Online Reinforcement Learning (AAAI 2024 · similarity 0.87)
- A Clean Slate for Offline Reinforcement Learning (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.
