Towards End-to-End Reinforcement Learning of Dialogue Agents for Information Access