A Bayesian Approach for Policy Learning from Trajectory Preference Queries