Gradient-based Optimization for Bayesian Preference Elicitation