Efficient Model-Free Reinforcement Learning Using Gaussian Process