General-to-specific learning of Horn clauses from positive examples

Abstract

The authors describe a method for learning disjunctive concepts represented as Horn clauses in a general-to-specific manner. They have identified a restricted class of Horn clauses for which positive examples are sufficient to detect overgeneral clauses. The method, developed and implemented in a system called INDICO, extracts as much constraining information as possible from the examples, such that the space of possible solutions can be searched efficiently. INDICO works in three steps. First, the argument types of the target predicate are determined. Second, the example set is partitioned and for each partition a clause head is determined which covers all the examples in the partition. Third, the clauses are specialized by adding literals, including newly invented ones, to their body until the definition is correct. Some experimental results are presented.<>