Research map / Reinforcement learning / Offline RL
Action-Free Offline-To-Online RL via Discretised State Policies
Natinael Solomon Neggatu, Jeremie Houssineau, Giovanni Montana
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 regularisation term to state-only DecQN (ablation) | OSO-DecQN · Hopper-medium-expert (action-free offline) | normalised return: 1.1 -> 110 | Improved |
| add regularisation term to state-only DecQN (ablation) | OSO-DecQN · Hopper-medium-replay (action-free offline) | discrete state difference error per step: 11.8 -> 0.99 | Improved |
| discretise state difference to Delta s instead of regressing next state s' (ablation) | behaviour cloning state policy · Hopper-medium-replay (action-free offline) | normalised return: 5.8 -> 29.2 | Improved |
| discretise state difference to Delta s instead of regressing continuous s'-s (ablation) | behaviour cloning state policy · Hopper-medium-replay (action-free offline) | normalised return: 4.9 -> 29.2 | Improved |
| discretise target to Delta s instead of s' (ablation) | behaviour cloning state policy · Quadruped-Walk-medium-expert (action-free offline) | discrete state difference error per step: 74.5 -> 14.7 | Improved |
| train with RL (OSO-DecQN) instead of imitating Delta s | BC Delta s · Halfcheetah-medium-expert (action-free offline) | normalised return: 54.7 -> 87.8 | Improved |
| train with RL (OSO-DecQN) instead of imitating Delta s | BC Delta s · Walker2d-medium-replay (action-free offline) | normalised return: 37.5 -> 84.8 | Improved |
| OSO-DecQN versus BC with actions on Walker2d-expert | OSO-DecQN · Walker2d-expert (action-free offline) | normalised return: 108.2 -> 108.1 | No clear effect |
| add guided online learning with pre-trained state policy | TD3 / DecQN_N · D4RL and DeepMind Control (online, 1M steps) | normalised return | Improved |
| add guided online learning on Hopper | TD3 · D4RL Hopper (online) | normalised return | No clear effect |
| compare Af-guide against TD3 baseline | TD3 · D4RL (online) | normalised return | Hurt |
| vary IDM architecture and batch size (ablation) | guided online learning · Online learning (appendix ablation) | normalised return | No clear effect |
Closest papers by meaning
- Offline Data Enhanced On-Policy Policy Gradient with Provable Guarantees (ICLR 2024 · similarity 0.91)
- Offline-Boosted Actor-Critic: Adaptively Blending Optimal Historical Behaviors in Deep Off-Policy RL (ICML 2024 · similarity 0.89)
- Beyond OOD State Actions: Supported Cross-Domain Offline Reinforcement Learning (AAAI 2024 · similarity 0.89)
- Offline RL with Observation Histories: Analyzing and Improving Sample Complexity (ICLR 2024 · similarity 0.88)
- Are Expressive Models Truly Necessary for Offline RL? (AAAI 2025 · similarity 0.88)
- A Clean Slate for Offline Reinforcement Learning (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.
