Why yet another software package?
---------------------------------

[Posted to OpenBayes by Kevin Murphy on 12 July 2001.]

I believe that the world definitely does not need yet another
program for handling toy Bayes nets. (By "toy" I mean: small,
discrete variables only, no learning, no choice of inference
methods, etc.) There is a large number of these already: see my list
at http://HTTP.CS.Berkeley.EDU/~murphyk/Bayes/bnsoft.html

There are some good packages available with a fair number of the features
above (e.g., Hugin, MIM, MSBN, Daphne Koller's code, Nir Friedman's
code, etc.), but none of these are open source (some are not even available
in executable form).

In addition, some of the existing libraries are hard to use. For
example, Koller's group has a C++ library called Phrog
(http://robotics.stanford.edu/~frog/), but several of her
students prefer to use BNT (from Berkeley!), because it is much
simpler to use. The lesson is that an interface in a high-level scripting
language is essential.

Hence there is still an unfulfilled need for a flexible, open-source library with
the features listed above. Although it will be impossible to satsify
the demands of all possible users (since each community has its own
language and needs), if we build a good collection of loosely-coupled tools
(in the tradition of Unix and Lego), people should be able to "roll their own"
applications with minimal effort.

The intended audience/user-base is broad: teachers, AI/stats
researchers, and people working on applications (e.g., biological
modelling, speech recognition, fault diagnosis, predicting the stock
market). By using default parameters and algorithms, it should be
possible to make the package easy to use for beginners, and yet
flexible enough for "power users".