Current approaches to privacy policy comparison use strict evaluation criteria (e.g. user preferences) and are unable to state how close a given policy is to fulfil these criteria. More flexible approaches for policy comparison is a prerequisite for a number of more advanced privacy services, e.g. improved privacy-enhanced search engines and automatic learning of privacy preferences. This paper describes the challenges related to policy comparison, and outlines what solutions are needed in order to meet these challenges in the context of preference learning privacy agents.