Gaussian Processes for Sample Efficient Reinforcement Learning with RMAX-Like Exploration