Reinforcement-learning systems chase rewards they are given, but nobody has found a reliable way to write one that captures ...