ameen@swivax.UUCP (Hanna) (02/21/89)
I'm interested in work concerning *model based* diagnosis in technical domains (electrical or mechanical). Especially *complex* systems (mostly *dynamic* in behavior) where appropriate abstract modeling should encounter this complexity. In particular, ideas/systems that reason about *multiple models* of the same device under diagnosis these include such issues as: o criteria for dynamic choosing of the current "best" model. (could be symptom directed). o criteria for determining that assumptions supporting the current considered model are no longer valid (e.g. diagnosis failed or inadequately discriminatory) and hence a different set of assumptions (implying a different model) should be tried. (One simple example is Davis' "bridge fault"). o non-monotonic reasoning about suspect components (a measure of belief that a certain suspect is indeed a culprit and further investigation in it's sub-components is worthwhile, it may indeed appear that this is not the case). o dealing with feedback loops among suspect components. o test cost considerations in suggesting test probes (other than just probability of failure like in DeKleer & Williams). Also heuristics to achieve the above mentioned goals and associative knowledge for diagnosis of "hard to model" components. A former work dealt with modeling the same device on different levels of behavioral abstractions (e.g. behavior of a chip output could be depicted in terms like: "pulse_train_generated", "frequency", "pulse_width" etc..) and each layer was built automatically from the level(s) underneath. Diagnosis, then, dealt with the highest abstract level which still provided a discrepancy. At certain levels a more refined level had to be considered. Each level (according to it's coarseness) possibly abstracted underlying structure). But the non-monotonic reasoning and explicit criteria for choosing the right level and the diversity of models concerning structure was lacking. Any ideas or references are appreciated, e-connection concerning model-based diagnosis is as well welcomed. Please e-mail to the e-address below. -- Ameen Abu-Hanna. (ameen@swivax.uucp). SPIN Project Social Science Informatics; University of Amsterdam Herengracht 196 Amsterdam 1016 The Netherlands.