Modelos probabilísticos para su utilización en sistemas expertos

Abstract

One of the most current problems with Expert Systems is the continuing argument between the defendants of probabilistic models and those against it. The alternative suggested (certainty factors, fuzzy logic, theory of evidence,� ) are inadequate at reproducing many real situations. Probabilistic models have been criticized either for the high number of parameters involved or the difficulties in their estimation from data. In fact, there is general agreement that the general dependence model is not practical for most real cases. Other sources of criticism come from the independence models, which are considered too simple for reproducing some real problems. As an alternative the model known as relevant symptoms dependence model is proposed. This model is not only theoretical but can be implemented as well, as shows the RSPS shell integrated in part of this thesis. Different systems of the representation of rules and the algorithms that allow changes from one representation to another are also analysed. Moreover some technical statistics that allow knowledge bases and inference engine to be created in expert systems are described. In relation to the learning of expert systems a method based on the maximum likelihood method for causal networks by Lauritzen and Spiegelhalter is proposed.