Near-optimal Optimistic Reinforcement Learning using Empirical Bernstein Inequalities