State Action Separable Reinforcement Learning