Data-Efficient Reinforcement Learning in Continuous-State POMDPs