Novel First Order Bayesian Optimization with an Application to Reinforcement Learning