Minimizing Live Experiments in Recommender Systems: User Simulation to Evaluate Preference Elicitation Policies