Could not download file: This paper is available to authorised users only.


M. El Alami1, M. Romero2, F. de Arriaga1

1Universidad Politécnica de Madrid (SPAIN)
2Universidad Nacional de Educación a Distancia (SPAIN)
The exponential growth of information and the reduced efficiency of the Internet browsers make imperative the solution of the filtering problem. The problem becomes more complex in the case of having a great number of users with different interest profiles.

The filtering problem implies a lot of decisions. Among them we quote: the determination of a user’s profile, the evaluation of the relevance or quality of a document according to that profile or the selection of filtering rules to attributes of the documents to be filtered, the adoption of a threshold for filtering, how much input requires the user, and who is going to use the filter: a single person or a group of people.

Probabilistic methods gave the first clue to solve the filtering problem but with less than satisfactory results; the main difficulty was the correct estimation of term-occurrence parameters. Starting in the late 80s, knowledge-based techniques have been extensively used. However, the problem of the filtering efficiency in any domain, by adapting dynamically the filtering to possible changes in the user’s interests is still open. It is the purpose of this paper.

The paper describes the efforts carried out to provide a considerable number of people with suitable documents retrieved from the Internet. We assume that users have very different interests in information. We want the system to be robust, flexible to adapt itself to possible changes in the needs of the users, and very friendly in the sense that the work done by the users to explain their search interests or to assess the system results has to be simple and easy.

For that purpose a multi-agent system has been implemented for this purpose. In order to provide learning capability to the intelligent agents, a set of flat functional-link neural networks which optimises the number of neurones in the network, the computer run-time, and the network training, has been devised. Besides, the system uses several fuzzy logic techniques to represent the degree of success of the system according to the user’s profile. The paper describes the different types of agents used by the system, the general architecture of the system and some of the fuzzy techniques implemented to solve the filtering problem. Some filtering results, automatically obtained by the system, are included to give an account of system precision.