Near-optimal Reinforcement Learning in Factored MDPs