Why EBL Produces Overly-Specific Knowledge: A Critique of the PRODIGY Approaches

Abstract There are many different ways to prove that a training example is subsumed by a target concept. Each proof gives rise to a different sufficient condition for the concept, some of which are considerably more general than others. Since EBL merely computes the weakest precondition of a particular proof, it is by no means guaranteed to find a sufficient condition that is maximally general. In practice, EBL frequently derives overly-specific control knowledge, retaining extraneous features of its training examples. In this paper we formally define the notion of a maximally general sufficient condition. We identify common pitfalls that prevent EBL from deriving such conditions, and critique the array of heuristic mechanisms used to improve EBL's generalizations in the PRODIGY system (including logical simplification, abstraction, and static analysis). Finally, we advocate the design of domain-independent, meta-level theories as a direction for future work.