MULLEN@SUMEX-AIM.ARPA (03/06/84)
From: Juanita Mullen <MULLEN@SUMEX-AIM.ARPA>
[Forwarded from the SIGLUNCH distribution by Laws@SRI-AI.]
Friday, March 9, 1984
LOCATION: Braun Lecture Hall (smaller), ground floor of Seeley Mudd
Chemistry Building (approx. 30 yards west of Gazebo)
12:05
SPEAKER: Ben Grosof
Stanford University, HPP
TOPIC: AN INEQUALITY PARADIGM FOR PROBABILISTIC KNOWLEDGE
Issues in Reasoning with Probabilistic Statements
BACKGROUND: Reasoning with probabilistic knowledge and evidence is
a key aspect of many AI systems. MYCIN and PROSPECTOR were pioneer
efforts but were limited and unsatisfactory in several ways. Recent
methods address many problems. The Maximum Entropy principle
(sometimes called Least Information) provides a new approach to
probabilities. The Dempster-Shafer theory of evidence provides a new
approach to confirmation and disconfirmation.
THE TALK: We begin by relating probabilistic statements to logic. We
then review the motivations and shortcomings of the MYCIN and
PROSPECTOR approaches. Maximum Entropy and Dempster-Shafer are
presented, and recent work using them is surveyed. (This is your big
chance to get up to date!) We generalize both to a paradigm of
inequality constraints on probabilities. This paradigm unifies the
heretofore divergent representations of probability and evidential
confirmation in a formally satisfactory way. Least commitment is
natural. The interval representation for probabilities includes in
effect a meta-level which allows explicit treatment of ignorance and
partial information, confidence and precision, and (in)dependence
assumptions. Using bounds facilitates reasoning ABOUT probabilities
and evidence. We extend the Dempster-Shafer theory significantly and
make an argument for its potential, both representationally and
computationally. Finally we list some open problems in reasoning with
probabilities.