Optimizing Dialogue Management with Reinforcement Learning: Experiments with the NJFun System