Comparing the Utility of State Features in Spoken Dialogue Using Reinforcement Learning