Offline RL

A reinforcement learning approach in which a policy is trained only on a fixed, pre-collected dataset without further interaction with the environment during training. Its main advantage is that training is reproducible and auditable, but it faces the theoretical difficulty of distribution shift.