Reinforcement Learning · advanced · concept 191 of 311 · Updated Nov 4, 2026
Offline Reinforcement Learning
Reinforcement learning from a fixed dataset of past interactions, with no new interaction while training. It matters where exploration is costly or unsafe: hospitals, robots, recommender logs. The core problem is distributional shift: a learned policy drifts toward actions the data never tried, and plain Q-learning overestimates their value. Conservative Q-Learning penalizes those actions; the Decision Transformer treats it as sequence modeling.
Top AI contributors on this
- Demis HassabisGoogle DeepMind (chair) / Isomorphic Labs
- Pieter AbbeelAmazon / UC Berkeley
- David SilverIneffable Intelligence / UCL
- John SchulmanThinking Machines Lab
- Sergey LevineUC Berkeley / Physical Intelligence
- Chelsea FinnStanford University / Physical Intelligence
Key terms
Learn these first
Where you meet it in the real world
Robot learning from logged trajectories, treatment policies from medical records, recommender and ad systems trained on logs, dialogue and LLM fine-tuning from static feedback data
Videos
Guides and articles
Hugging Face
Courses, papers, and more