I am pleased to announce a major new release of the Bayes Net Toolbox, a software package for Matlab 5 that supports inference and learning in directed graphical models. Specifically, it supports exact and approximate inference, discrete and continuous variables, static and dynamic networks, and parameter and structure learning. Hence it can handle a large number of popular statistical models, such as the following: PCA/factor analysis, logistic regression, hierarchical mixtures of experts, QMR, DBNs, factorial HMMs, switching Kalman filters, etc. For more details, and to download the software, please go to http://www.cs.berkeley.edu/~murphyk/Bayes/bnt.html The new version (2.0) has been completely rewritten, making it much easier to read, use and extend. It is also somewhat faster. The main change is that I now make extensive use of objects. (I used to use structs, and a dispatch mechanism based on the type-tag system in Abelson and Sussman.) In addition, each inference algorithm (junction tree, sampling, loopy belief propagation, etc.) is now an object. This makes the code and documentation much more modular. It also makes it easier to add special-case algorithms, and to combine algorithms in novel ways (e.g., combining sampling and exact inference). I have gone to great lengths to make the source code readable, so it should prove an invaluable teaching tool. In addition, I am hoping that people will contribute algorithms to the toolbox, in the spirit of the open source movement. Kevin Murphy
