Augmenting Domain Theory for Explanation-Based Generalization

Publisher Summary This chapter addresses one of the shortcomings of explanation-based generalization (EBG), namely, that EBG requires the domain theory to be complete. EBG is a form of learning that uses a strong domain theory to constrain the search for generalizations of the concept that is being learned. EBG attempts to explain why the example is a positive instance of the target concept. As the domain theory already has general rules for proving the concept, the benefit of the method is its ability to operationalize and generalize the explanation. The system is capable of learning disjunctive structured concepts one disjunct at a time. It can acquire new rules and missing disjuncts, and cope with an expanding instance language. It also forms operationalized definitions of the concept to improve performance. It learns best when several concepts with shared subconcepts are learned simultaneously. Disadvantages of this approach comprise an inability to learn intermediate concepts and missing conjuncts. It also forms overly specific rules in contexts where one conjunct is invariant. The system is not discriminating enough when linking predicates to find new rules. Some form of background knowledge as in OCCAM (Pazzani, 1988) would be useful. An ability to detect and retract erroneous rules would also be useful. More discriminating forms of generalization would also allow learning of more complicated concepts.