Learning Cartesian Granule Feature Models

Abstract

In the previous chapter, it was shown how Cartesian granule feature models exploit a divide-and-conquer strategy to representation, capturing knowledge in terms of a network of low-order semantically related features. Both classification and prediction problems can be modelled quite naturally in terms of these models. This chapter describes a constructive induction algorithm, G_DACG (Genetic Discovery of Additive Cartesian Granule feature models), which facilitates the learning of such models from example data [Shanahan 1998; Shanahan, Baldwin and Martin 1999]. This involves two main steps: language identification (identification of the low-order semantically related features in terms of Cartesian granule features); and parameter identification of class fuzzy sets and rules. The C_DACG algorithm achieves this by embracing the synergistic spirit of soft computing, using genetic programming to discover the language (structure) of the model fuzzy sets and evidential rules for knowledge representation, while relying on the well-developed probability theory for learning the parameters of the model.