I think you'll find the following list much more up-to-date
http://HTTP.CS.Berkeley.EDU/~murphyk/Bayes/bnsoft.html
(Russell Almond has not changed his page since 1996, and told me that he
has no plans to do so.)
Rich told me he wants to learn structure, not (just) parameters.
According to my list, there are 3 free packages that can do this:
"Bayesian Knowledge Discoverer" from Open Univ., "Belief Net Power
Constructor" from Ulster Univ., and my "Bayes Net Toolbox".
The first uses "bound and collapse" to do inference with missing data
(instead of the better known EM algorithm), the second cannot handle
missing data, and the third only learns the inter-slice connectivity
matrix of a DBN using structural EM.
So, Rich, it sounds like BKD is the one to use.
http://kmi.open.ac.uk/projects/bkd/
HTH,
Kevin
