Technical University of Munich, Department of Computer
MQM Paper: [Struss 92d]
Diagnosis as a Process
In: Readings in Model-based Diagnosis. Eds.: Walter Hamscher, Luca Console, Johan de Kleer. Morgan Kaufmann Publishers, pp. 408-418, 1992.
The paper proposes to make efforts in order to develop a theory of diagnosis that
facilitates exchange of ideas and experience among researchers in the field.
The key idea is to explicitly represent assumptions and choices that drive the diagnostic process, just as assumptions about the appropriateness of parts of the model are recorded, and to "diagnose" them in the same manner. Such working assumptions may concern, for instance, the single fault hypothesis, the completeness of knowledge about fault modes, or the absence of intermittent faults. Knowledge about situations that suggest questioning and eventually abandoning such hypotheses constitutes an important part of diagnostic skills. Because this kind of knowledge will basically consist of experience and heuristics, the work is also guided by the aim of deriving an appropriately structured interface between model-based and heuristic reasoning in diagnostic systems.
Current model-based diagnostic systems or theories mainly concentrate on the aspect of deriving information about the correctness of components from discrepancies between observations (measurements) and predictions based on a device model ([de Kleer- Williams 87], [Reiter 87], [DagueDeves-Raiman 87], [Hamscher 88], [Struss 88a,b], [de Kleer-Williams 89], [Raiman 89]). From this perspective, the evolution of the diagnostic process is basically regarded as a sequence of measurements in a given situation, and determining the best next measurement is considered to be its essential task (for instance, based on a probabilistic theory as in the General Diagnostic Engine, GDE, of [de KleerWilliams 87], or in [Hamscher 88]). In [Struss 88a,b], it was demonstrated how other types of diagnostic reasoning can be integrated into such a framework: focusing in hierarchical models, performing multiple tests and interpreting and analyzing observations. Only a sketch was given of how other subtasks and knowledge sources could be incorporated which might occur in a concrete diagnostic task.
This paper elaborates this point and attempts a step towards a theory that, in a very general way, treats diagnosis as a complex, non-monotonic process. Such a theory should provide foundations for diagnostic systems that are both more general and more specific compared to the frameworks proposed so far:
Although it may not be a basis for efficient application systems, such a generic diagnostic system can serve as a flexible basis for experimenting with different modeling techniques and diagnostic strategies. Despite of the requirement for a new architecture, this can be based on crucial elements of GDE, namely model- based prediction and ATMS-based reasoning.