Active Learning of Robot Reward Functions