GART: a tool for experimenting with approximate reasoning models

摘要

This paper presents GART, a software tool designed for carrying out experimental activities with approximate reasoning models. GART has been built on the basis of a generic conceptual model of approximate reasoning, which encompasses both fuzziness and uncertainty. The main dimensions of such a model have been analyzed and all the values they can assume have been considered. The flexibility of the conceptual model underlying GART is substantiated by a parametric object-oriented data structure, which accommodates the most significant design choices of a non-classical reasoning theory. GART can support the designer of an approximate reasoning system in identifying the most suitable reasoning model for the application at hand, and in performing a tuning of the features characterizing the reasoning mechanism. Besides, GART can support the AI researcher. In fact, an extensive experimentation against a set of reasoning cases from different domains can be carried out with the goal of establishing associations between domain characteristics and the features of the most appropriate approximate reasoning model for a given application context. Experimenting with GART is made easier by a graphical interface, which allows the user to define the features of an approximate reasoning model and to develop a test knowledge base for a specific context. GART has been validated using case studies from three different application domains: preventive diagnosis of power transformers, diagnosis of acute coronaric ischemia, and classification of historical violins.