Decoupling Exploration and Exploitation in Reinforcement Learning