  |
|
This is a development site
and this page is still under construction.
PURSUIVANT
 ink from
here to the PURSUIVANT
Home Page at the Univerity of Kansas. The PURSUIVANT
project is currently supported by the Ernst & Young
Center for Auditing Research and Advanced Technology (E
& Y CARAT) at the KU School of Business under the
direction of Dr. Rajendra P. Srivastava.
The project is using Delpi to
develop PC-based software to facilitate the
implementation of Valuation Networks. Programming is
being carried out by computer science undergraduate
students from KU, and project management is the
responsibility of Dr. Peter R. Gillett at Rutgers.
Based on the work of Glenn Shafer
and Prakash Shenoy, Valuation Networks began as an
axiomatic generalization of probability theory and
Dempster-Shafer belief functions sufficient for
propagation using local computation in a join tree.
In the current version of the
software, users model domain problems (e.g., audit
planning applications) by using a graphical interface to
construct a network of variables and related valuations;
valutions may be instantiated as probability potentials,
belief-function potentials, Sphonian epistemic
(dis)belief potentials, possibility potentials, or
propositional truth values. PURSUIVANT is
designed to construct a join tree from the user network
in the form of a Shenoy-Shafer join tree, a binary join
tree, a junction tree, or a junction tree with
separators. Valuations are then propagated using the
Shenoy-Shafer or the Aalborg methods and appropriate
rules for combination, marginalization and normalization
of the chosen calculus. Alternative heuristics for
constructing the trees (e.g., one-step look ahead) are
being programmed.
Later versions of the software will
permit users to define combination, marginalization and
normalization rules for new calcui for uncertain
reasoning that obey the Shenoy-Shafer axioms, and will
also provide decision analysis features. The current
softtware will display the generated join tree and will
provide computed output marginals. In the case of belief
functions, valuations may be input or displayed in any of
the usual four representations (basic probability
assignments, belief functions, plausibility functions, or
commonalities): fast Moebius transforms are used to
convert these as required.
From here you may wish to re-vist my Home Page, return to the Departure
Board to visit another part of the site, or leave via
the Exit.
Last Updated: 08/26/09
Dr. Peter R. Gillett
| Thank you for
visiting my Web Site. Please visit again soon! |
|
|
