Interactive Learning from Policy-Dependent Human Feedback