Offline reinforcement learning is a model learning a policy from a fixed, previously collected set of records without interacting with the environment at all. In standard reinforcement learning the agent tries something, sees the result and adjusts. Here there is no trying. All that exists is decisions taken in the past and what followed them.
In some domains that constraint is not optional. In healthcare, finance, industrial control and autonomous driving, letting a model experiment on live systems is unacceptable because a wrong move costs something real. In those settings the historical operational logs become the dataset and learning happens on top of them.
The core technical problem is distribution shift. The model can predict that an action never tried in the dataset would work brilliantly, with no record anywhere to check that prediction against. Algorithms therefore push the policy to stay near behaviour actually seen in the data and penalise over-optimistic values for unobserved actions.
A concrete case: a call centre holds recordings and transcripts of 400,000 past calls, including when a discount was offered, when a call was escalated and how each ended. Offline learning derives a routing policy from those records, then gets tested against a separate period of logs before going anywhere near live traffic.
Quality depends entirely on how varied the records are. If the same decision was always made in the past, the model cannot learn whether the alternative would have been better.