How agents should learn from human preference