How to Build User Simulators to Train RL-based Dialog Systems