Online Model Selection for Reinforcement Learning with Function Approximation