Possible Thesis Topics for Qualified Students
Comparison of Existing Shells for Bayesian Networks
and Influence Diagrams
This is an M.S. (P) thesis topic.
It may also be a good topic for an honors thesis.
The student would design a set of evaluation
criteria (covering, e.g., user interface issues, functionality, and
efficiency)
to compare existing shells and evaluate, at least, the ones that are
available for academic use free of charge, including
Hugin Light (www.hugin.dk), MSBN (research.microsoft.com/dtas/msbn), GeNie
(www2.sis.pitt.edu/~genie), Netica (www.norsys.com), and WEBWEAR-III
(cs.uregina.ca/~yxiang/ww3).
Evaluation will include some experimentation, for example tied to a
500-level course.
CB, Revisited---Completed by Bing Xia (M.S. Thesis)
This is an M.S. (P) thesis topic.
The student would implement the Bayesian network learning algorithm
CB with a polished user interface.
The implementation would be sufficiently
robust and efficient for use by non-specialists and sufficiently modular and
well documented to be extended and integrated with other Bayesian network
software tools.
SQL Implementation of the Junction Tree Method for Probability Update in
Bayesian Networks (Close to Completion: Jincao Ye's (R) M.S. Thesis)
This is an M.S. (P) thesis topic.
The student would implement the Junction Tree Method in SQL. This work would
be based on algorithms by Wong, Butz, and Xiang, as published in the
UAI-95 proceedings. The implementation would run on the NCR Teradata computer
recently acquired by Professors Huhns and Vargas for USC.
The implementation would be sufficiently
robust and efficient for use by non-specialists and sufficiently modular and
well documented to be extended and integrated with other Bayesian network
software tools.
The Factor Tree Method for Probability Update in Bayesian Networks
This is an M.S. (P or R) thesis topic. It may grow into a Ph.D. topic.
The student would implement the factor tree method of Bloemeke (Bloemeke and
Valtorta, UAI-98) and some of its variants.
Time and space efficiency would be more important than usability,
extensibility, and integrability.
Experiments would be run to compare the efficiency of probability update using
factor trees to other exact methods (e.g, junction trees, SPI).
Learning Bayesian Networks from Relational Databases
This is an M.S. (R) topic. It may grow into a Ph.D. topic.
The student would study how to exploit the structure of relational databases
for Bayesian network learning, propose algorithms to carry out this task, study
their complexity, and carry out a research-style implementation. (A starting
point may be Friedman, Getoor, Koller, and Pfeffer's "Learning Probabilistic
Relational Models," IJCAI-99.)
Beginnings of a Causal Modeler's Workbench
This is an M.S. (P) thesis topic. It may grow into a Ph.D. topic.
The student would implement a basic shell for causal Bayesian networks,
supporting the criteria for identifiability, such as the back door criterion,
as described by Pearl (Biometrika, 1995; Causality, Chapter 3,
2000), and extending the C implementation of the back door criterion
by Woody Folsom. The shell would include a polished user interface.
The implementation would be sufficiently
robust and efficient for use by non-specialists and sufficiently modular and
well documented to be extended and integrated with other Bayesian network
software tools.
Soft Evidential Update
This is an M.S. (P) thesis topic. It may grow into a Ph.D. topic.
The student would implement the two algorithms for soft evidential update by
Valtorta, Vomlel and Kim (2000, submitted).
Time and space efficiency, usability,
extensibility, and integrability would all be important criteria.
Evidence in Decision Networks
This is an M.S. (R) thesis topic. It may grow into a Ph.D. topic.
The student would explore the meaning of evidence, both hard and soft, in
influence diagrams, design a modified version of the strong junction tree
algorithm that correctly deals with influence, and implement the algorithm.