Formalizing elaboration tolerance
John McCarthy, Aarati Parmar · 2003
This thesis examines the concept of elaboration tolerance. A representation, whether it be a logical theory, computer program, or even Bayesian network, is elaboration tolerant to the extent it is easy to change to represent new facts. Natural language is the model for elaboration tolerance; in everyday discourse one can easily and succinctly change one's declarations by adding an extra, sentence or two. This thesis has two major contributions. The first is a framework which elucidates how the semantics of a representation affects its elaboration tolerance, and the viability of principles that promote elaboration tolerance. The second contribution is a thorough examination of what we call additive elaboration tolerance. In this paradigm, a representation is modified simply by adding the formula whose meaning corresponds to the desired change. This modality of changing representations is attractive because it avoids “brain surgery”—we do not have to tinker explicitly with the representation at all. We show how to endow a given representation S1 with additive elaboration tolerance by embedding it in another system SAb · S 1 = S1,Ab. This embedding is intuitive as it models human discourse. It is also sound, preserving facts true in tire original representation. Elaboration tolerance shares concepts with the fields of belief revision and reformulation. We explore how the AGM postulates relate to our construction S1, Ab. We also show how standard problems of nonmonotonic reasoning can be re-framed in terms of searching for elaboration tolerant representations of commonsense theories. We conclude with deeper insights about elaboration tolerance, and its importance to the future of the logical AI program.