Accommodating Picky Customers: Regret Bound and Exploration Complexity for Multi-Objective Reinforcement Learning