Due to the astonishing speed at which new content is created and published on the Web, it is increasingly difficult for users to make the most appropriate decisions in front of an overwhelming amount of information. Recommender systems try to help users by analyzing and ranking the available alternatives according to their preferences and interests, modeled in user profiles. One important problem to solve in the development of these systems is how to discover the user preferences, and how to maintain them dynamically. In this work we propose to use the information given by a user in his/her interaction with the recommender system (e.g. the selection of the news to be read every morning) to infer his/her preferences on several criteria on which the decision alternatives are defined. More specifically, the paper is focused in learning the most preferred value for the user in the case of numerical attributes. A methodology to adapt the user profile in a dynamic and automatic way is presented. The adaptations may be performed after each interaction of the user or after the system has gathered enough information from several user selections. We have developed a framework for the automatic evaluation of the performance of the adaptation algorithm that permits to analyze the influence of different parameters. The obtained results show that the adaptation algorithm is able to learn a very accurate model of the user preferences after a certain amount of interactions.