Inference without Chaining

Abstract

The problem of specifying, constructing , and understanding specialized, limited inference systems arises in many areas of AI. As a first step towards solving this problem this paper recommends the development of an inference engine that is limited by its inability to chain together two pieces of a representation in order to derive a third. A method for using model theory to specify limited inference is introduced and then used to specify an inference engine via a t hree-valued logic. This inference engine is proved to be the strongest one that does no chaining, modulo the way that it divides the representation into pieces. Thus , the specification captures the set of all inferences that require no chaining. This paper also surveys and compares a number of systems that do no chaining as well as some that allow only selected forms of chaining.