Abductive diagnosis using time-objects: criteria for the evaluation of solutions
Journal
Computational Intelligence
Date Issued
September 17, 2002
Author(s)
DOI
10.1111/0824-7935.00135
Abstract
Diagnostic problem solving aims to account for, or explain, a malfunction of a system (human or other). Any plausible potential diagnostic solution must satisfy some minimum criteria relevant to the application. Often there will be several plausible solutions, and further criteria will be required to select the "best" explanation. Expert diagnosticians may employ different, complex criteria at different stages of their reasoning. These criteria may be combinations of some more primitive criteria, which therefore should be represented separately and explicitly to permit their flexible and transparent combined usage. In diagnostic reasoning there is a tight coupling between the formation of potential solutions and their evaluation. This is the essence of abductive reasoning. This article presents an abductive framework for diagnostic problem solving. Time-objects, an association of a property and an existence, are used as the representation formalism and a number of primitive, general evaluation criteria into which time has been integrated are defined. Each criterion provides an intuitive yardstick for evaluating the space of potential solutions. The criteria can be combined as appropriate for particular applications to define plausible and best explanations. The central principle is that when time is diagnostically significant, it should be modeled explicitly to enable a more accurate formulation and evaluation of diagnostic solutions. The integration of time and primitive evaluation criteria is illustrated through the Skeletal Dysplasias Diagnostician (SDD) system, a diagnostic expert system for a real-life medical domain. SDD's notions of plausible and best explanation are reviewed so as to show the difficulties in formalizing such notions. Although we illustrate our work by medical problems, it has been motivated by consideration of problems in a number of other domains (fermentation monitoring, air and ground traffic control, power distribution) and is intended to be of wide applicability