Global Policy Construction in Modular Reinforcement Learning