A new system for inductive learning in attribute-based spaces

Abstract

Despite the fact that the theory and methodology of inductive learning has previously been described, actual symbolic systems for learning descriptions in attribute-based spaces use other algorithms. The AQ family of programs uses properly guided logic-based negation, union, and intersection operators. The ID family of programs uses the idea of iteratively partitioning the search space. The inductive methodology, on the other hand, describes a number of generalization and specialization operators, some of which can be restricted to the attribute-based spaces. Why, then, they are not used in practice? The reasons are the immense search spaces of inductive learning and its weak heuristics.

This paper describes an approach that uses such task-specific inference rules directly on formulas of a rule-based framework. To control such a search, in the absence of strong heuristics, we use the very liberal and robust search mechanism of genetic algorithms. A simple implementation is tested on two standards data sets. The results suggest the feasibility of this approach, not only in terms of numerical qualities, but also in terms of understanding the output knowledge and the processing mechanisms.