Reference: Levy, A. Exploiting (Ir)relevance to Guide Problem-solving. Knowledge Systems Laboratory, May, 1991.
Abstract: Controlling inference in Artificial Intelligence (AI) systems that use declarative representations has always been a key issue in AI research. The problem arises because the problem-solver's search is unfocused and considers many irrelevant facts. In the past, meta-level reasoning and creating abstraction hierarchies have tried to deal with this problem. The notion of irrelevance plays an important role in both fields. A significant type of meta-level control knowledge are statements about irrelevance of entities in the representation to problem-solving goals. Given a goal, it is possible to state that some entities in our KB are irrelevant to it and should therefore not be used in the search for its solution. In the case of abstractions, defining the abstraction hierarchy is equivalent to stating what aspects of the domain are relevant at each level. I describe a framework in which meta- level control facts about (ir)relevance can be stated and subsequently used by a problem-solver. These control facts, called relevance-claims are stated in a declarative fashion, much the same way as the knowledge about the domain. The claims state relevance between problem-solving goals and entities in the KB. Their meaning is defined by the way they constrain the possible space of deductions that the problem-solver considers. The framework will be studied and implemented on a device-modeling system and on the Cyc system.