Learning Trajectory Preferences for Manipulators via Iterative Improvement