> if my Belief Networks starts with unprecise prior probabilities, can I get
> precise results? How and Why? Any advices are highly appreciated.
I'm not sure the problem has been investigated as stated, but one place to
look at is the following paper:
@ARTICLE{pradhan96,
AUTHOR = "Pradhan, Malcolm and
Henrion, Max and
Provan, Gregory and
del~Favero, Brendan and
Huang, Kurt",
TITLE = "The Sensitivity of Belief Networks to Imprecise
Probabilities: An Experimental Investigation",
JOURNAL = "Artificial Intelligence",
VOLUME = 85,
NUMBER = "1--2",
MONTH = AUG,
YEAR = 1996,
PAGES = "363--397"
}
The paper shows experimentally that there are good reasons not to worry too
much about precision of numerical parameters in Bayesian networks.
Other directions to look at are: sensitivity analysis (finding out what
matters in a model and increasing precision of those parameters; in
addition to classican decision analysis, there was some work published in
UAI, some names here are: Laskey, Cozman), expected value of reducing
uncertainty (Henrion's dissertation and book "Uncertainty").
Marek
--------------------------------------------------------------------------
Marek J. Druzdzel http://www.pitt.edu/~druzdzel