How well a learned model predicts environment behavior for state-action combinations beyond those in its training data.