A statistical technique to separate signal (action effects) from noise (randomness in later steps) in reward observations.