Efficient Reinforcement Learning using Gaussian Processes