NAME:  			John D. Lowrance

Thesis year:		1982

Thesis title:		Dependency-Graph Models of Evidential Support

Thesis abstract - BEGIN

	Dependency-graph models of evidential support are formal systems
capable of pooling and extending evidential information, while maintaining
internal consistency. In this formalism, the likelihood of a proposition is
represented as a subinterval of the until interval.  The lower bound
represents the degree of "support" provided a proposition by a body of
evidence, and the upper bound represents the extent to which it remains
"plausible."  The smaller this interval, the more precisely the probability
of that proposition is known.
	Evidential information, extracted from the environment by
(indivisible) sources of knowledge, enters these models in the form of
probability "mass" distributions, defined over sets of propositions common
to both them and the model.  Theses mass distributions are combined through
Dempster's rule of combination [Dempster 1967].  The result is a new mass
distribution representing their consensus.  Next, this pooled information
is extended from those propositions it directly bears upon, to those it
indirectly bears upon, and converted to the interval representation.  Prior
probabilities, frequently difficult or impossible to collect in artificial
intelligence domains, but required by most other systems of inexact
reasoning, are not needed. This form of evidential reasoning, based on
[Shafer 1976], is more general than either a Boolean or Bayesian approach,
providing for Boolean and Bayesian inferencing when the appropriate
information is available.
	Dependency graphs are formal representations of dependency
relations. A dependency graph consists of a set of propositions (nodes), a
covering assignment of confidences (node values), and a coordinated set of
dependency relationships (connecting arcs) constraining the assignment of
confidences.  Confidences can be fully specified (a single value),
partially specified (several values), or unspecified (all values).
Similarly, a dependency graph can describe any degree of
dependence/independence among its propositions.  The freedom to express
partial information makes dependency graphs suitable for modeling the
degrees of belief one should accord a group of related propositions based
on evidential information, a suitable host for Shafer's theory.

Thesis abstract - END

Department:			Computer and Information Science (COINS)

University:			University of Massachusetts

Univ. location:			Amherst, MA

Thesis advisor:			Edward M. Riseman
Advisor's department:		COINS

Committee member:		Michael A. Arbib
Member's department:		COINS

Committee member:	      	Daniel P. Friedman
Member's department:		Computer Science, Indiana University

Committee member:		Victor R. Lesser
Member's department:		COINS

Research group:			Artificial Intelligent Center
Research institution:		SRI International
Institution location:		Menlo Park, CA
Dates:				9/29/80 - present