Model-Free Preference-Based Reinforcement Learning