voula@utcsri.UUCP (Voula Vanneli) (04/10/85)
UNIVERSITY OF TORONTO
DEPARTMENT OF COMPUTER SCIENCE
(SF = Sandford Fleming Building, 10 King's College Rd.)
ARTIFICAL INTELLIGENCE SEMINAR - Monday, April 15, 11 a.m.,
SF 1105
Harry Stephanou
Texas
"Processing Uncertain Knowledge"
Processing Uncertain Knowledge
Abstract
The methodology described in this talk is motivated by
the need to design knowledge based systems for applications
that involve: (1) subjective and/or incomplete knowledge
contributed by multiple domain experts, and (ii) inaccurate
and/or incomplete data collected from different measure-
ments. The talk consists of two parts.
In the first part, we present a quantitative criterion
for measuring the effectiveness of the consensus obtained by
pooling evidence from two knowledge sources. After a brief
review of the Dempster-Shafer theory of evidence, we intro-
duce a set-theoretic generalization of entropy. We then
prove that the pooling of evidence by Dempster's rule
decreases the total entropy of the sources, and therefore
focuses their knowledge.
In the second part of the talk, we present a fuzzy
classification algorithm that can utilize a limited number
of unreliable training samples, or prototypes. We then pro-
pose the extension of this algorithm to a reasoning by anal-
ogy scheme in which decisions are based on the "similarity"
of the observed evidence to prototypical situations stored
in the knowledge base. The measure of similarity relies on
a set- theoretic generalization of cross-entropy.