Reinforcement Learning in POMDPs with Function Approximation