Extrapolation Error

In offline reinforcement learning, the error in value function estimation that arises when a policy selects actions not covered by the dataset; the error accumulates as trajectories are rolled out.