How Does Human Reinforcement Learning Cope with Varying Task-Irrelevant Features?