Hello all,

I am happy to say that my dissertation is available for your
reading pleasure -- a misspelling of ``Parmenidean'' has been
corrected. :)  You will find .ps, .ps.gz, and .pdf versions
on my publications page,

  http://civil.colorado.edu/~dodier/publications.html

>From the point of view of UAI folks, I believe the features
of greatest interest are these:

(1) A scheme for handling a wider range of conditional 
distributions -- an attempt is made at run time to determine
whether an exact result can be computed and otherwise an
approximation is made. Toward this end a catalog of exact
results has been accumulated (Appendix C). Approximations are
of two sorts -- the mixture of Gaussians which is doubtless
familiar to you, and monotone cubic splines, which are perhaps
a little less familiar; I've become a big fan of monotone
cubic splines.

(2) An implementation of distributed belief networks. Belief
networks can be spread across multiple Internet hosts and
communicate by a message-passing algorithm. Parent variables
can be named by host and a belief network on that host. Each
host runs software to carry out computations locally and to
exchange messages with parents and children on other hosts.
I used Java's Remote Method Invocation classes to implement
the message-passing; at the bottom are socket connections.

I have found that even if all belief networks are running on
a single host, a decomposition of a model into a distributed
belief network (rather than a monolithic one) can speed model
development and make modifications easier.

(3) Engineering applications are presented which include
sensor fusion, diagnosis of equipment status, assessment of
the influence of model variables by mutual information, and 
analysis of uncertainty in a first-principles model. 

I have found distributed belief networks to be very effective
models in engineering problems, because it is so natural to
model different sources of information and fuse them together
using the laws of probability. I look forward to further 
progress in this field!

Enjoy,
Robert Dodier

Reply via email to