Received: (from rga6711@localhost) by comet.research.ets.org (8.9.1b+Sun/8.9.1)
I'm working on an algorithm for learning the structure of a bayesian
network from data using a bayesian approach. In order to evaluate my
algorithm, I'm looking for a free and simple program that create a database
of cases from a given structure. Once the database detected, I will run my
algorithm on it and compare the initial structure with the structure detected.
The algorithim is so simple that any Bayes net propagation engine could in
principle do it. (If it comes as a shared library, it is relatively
simple to write a wrapper program to do this. I did it recently using
the Ergo DLL.) First, select a node and get its marginal
distribution, randomly select a value according to that distribution.
Next, set that node to the randomly selected value and propagate. Now
choose another node, and find its marginal distribution (conditioned
on the value of the first node). Randomly choose its value according
to that conditional distribution. When you have instantiated all of
the nodes, you will have sampled a value from this Bayes net.
Can anyone save me the time and recommend me a link, an adress or a person
working on such programs?
Free means you need to do work yourself to get it going. There are a
couple of options.
a) BELIEF had this capbility. It is available in the CMU AI
reposititory, but it is written in a rather old (pre-ANSI) version of
Common Lisp so it needs a few fix-ups to work in a modern Lisp.
<a
href="http://www.cs.cmu.edu/afs/cs/project/ai-repository/ai/areas/reasonng/probabl/belief/0.html">CMU
Artificial Intelligence Library</a>
b) BUGS will also sample from an arbitrary graphical model (includinge
Bayes nets). However, it will take you a bit of effort to learn to
use it and describe your model in BUGS. Bugs is at:
<a href="http://www.mrc-bsu.cam.ac.uk/bugs">England</a>
<a href="ftp://muskie.biostat.umn.edu/mirror/methodology/bugs">US Mirror</a>
c) Joe Schafer the Penn State Statistics Department has some programs
for multiple imputation in missing data that might be pressed into
service. His programs cover multivariate-multinomial (Bayes nets),
multivariate normal and conditional Gaussian models.
Good luck.
--Russell Almond
Educational Testing Service
Research Statistics Group, 15-T
Princeton, NJ 08541
Phone: 609-734-1557 FAX: 609-734-5420
Email: [EMAIL PROTECTED], [EMAIL PROTECTED]
http://www.stat.washington.edu/bayes/almond/almond.html
[Remove -- from email addresses]
