Synthesizing Neural Network Controllers with Probabilistic Model based Reinforcement Learning