dsn%umcp-cs.csnet@csnet-relay.arpa (03/27/84)
From: Dana S. Nau <dsn%umcp-cs.csnet@csnet-relay.arpa>
From: Toby Robison <eosp1!robison>
I think it is hopeless to demand that the algorithms instanced by
expert systems be well understood so they can be questioned. Even
when the algorithms can easily be printed, they will be hard for any
human being to comprehend, except in the most trivial systems. ...
I disagree. One of the reasons for separating an expert system's control
structure from the knowledge base is to allow for complex behavior with a
simple control algorithm. For example, Mycin's control structure is only
about one typewritten page [1]. Jim Reggia and I at the Univ. of Maryland
are currently working on a considerably more complex expert system control
structure, but even it is not THAT hard to understand once one understands
the preliminary mathematical background [2]. We even have a proof of
correctness for the algorithm!
REFERENCES:
[1] Davis, Buchanan, and Shortliffe. Production Rules as a Representation
for a Knowledge-Based Consultation Program. ARTIFICIAL INTELLIGENCE 8
(1977), 15-45.
[2] Reggia, Nau, and Wang. A Theory of Abductive Inference in Diagnostic
Expert Systems. TR-1338, Computer Sci. Dept., Univ. of Maryland (Dec.
1983). Submitted for Publication.borgia@cwruecmp.UUCP (The Prince) (04/04/84)
Don't we always come back to the same old epistemological question? What good is any system to a human being unless it is understood (may be in parts) by one or more human beings? An expert system should be an intelligent and articulate student who learns from several experts. Understanding control structures is not the critical issue. Well-known and fairly simple inference mechanisms are available. The key issue is articulating what knowledge was used and how in solving a problem. "What good is knowledge when it brings no profit to its bearer" - Teireisias in Oedipus the King, Sophocles -- joe borgia usenet: decvax!cwruecmp!borgia csnet: borgia@case arpanet: borgia.case@csnet-relay