Are Substitutions the Better Examples? Learning complete Set of Clauses with Frog.

The paper presents an approach for machine learning in a restricted first-order language with finite minimal Herbrand models by means of a search through a propositional representation space. The learning target is to find a set of goal clauses which can be used to define a target predicate. That is, we deal with single-predicate learning. For the search process we use the learning algorithm JoJo/ Frog which provides a flexible search strategy. The transition from the first-order representation to the representation in propositional logic is achieved by ground substitutions which transform clauses into ground clauses. Taking a closer look at this transition makes clear that the sufficiency condition which is used by algorithms like FOIL and LINUS as a criterion for judging the achieved learning results does not correspond to the completeness condition in the propositional case. Therefore, we use an extended completeness condition which captures all information given by the example knowledge. As a consequence we get a new definition of positive and negative examples. Instead of ground facts we regard ground substitutions as examples.