Inferring a reward function from observed behavior or human feedback rather than specifying it directly.