FAWCETT@RED.RUTGERS.EDU.UUCP (09/09/86)
Next week, on Tuesday, September 16th in Hill 423 at 11 AM, Jack
Mostow will give a talk based on his work with Kerstin Voigt, entitled
"A Case Study of Goal Integration in Heuristic Algorithm Design".
This a joint ML/III seminar, and is a dry run for a talk being given at the
Knowledge Compilation Workshop. There's no paper for the talk, but Jack
recommends his AAAI86 article with Bill Swartout as good background reading.
The abstract follows:
Jack Mostow
Rutgers University
(Arpanet address: MOSTOW@RED.RUTGERS.EDU)
A Case Study of Goal Integration in Heuristic Algorithm Design:
A Transformational Rederivation of MYCIN's Therapy Selection Algorithm
An important but little-studied aspect of compiling knowledge into
efficient procedures has to do with integrating multiple, sometimes
conflicting goals expressed as part of that knowledge. We are
developing an artificial intelligence model of heuristic algorithm
design that makes explicit the interactions among multiple goals. The
model will represent intermediate states and goals in the design
process, transformations that get from one state to the next, and
control mechanisms that govern the selection of which transformation
to apply next. It will explicitly model the multiple goals that
motivate and are affected by each design choice.
We are currently testing and refining the model by using it to explain
the design of the algorithm used for therapy selection in the medical
expert system MYCIN. Previously we analyzed how this algorithm
derives from the informal specification "Find the set of drugs that
best satisfies the medical goals of maximizing effectiveness,
minimizing number of drugs, giving priority to treating likelier
organisms, [etcetera]." The reformulation and integration of these
goals is discussed in Mostow & Swartout's AAAI86 paper. Doctoral
student Kerstin Voigt is implementing a complete derivation that will
address additional goals important in the design of the algorithm,
such as efficient use of time, space, and experts.
-------