Announcement: Bayesian Networks in Java (BNJ) community

We have recently created a LiveJournal community, bayesnets, for discussion of Bayesian Networks in Java, the most comprehensive suite of open-source Java tools for inference and learning using graphical models.

BNJ is developed using Eclipse and Borland JBuilder, and is distributed under the GNU General Public License (GPL).



BNJ consists of modules that implement:

• ConverterFactory, a GUI-based application for format conversion among Bayesian network formats such as Hugin (.hugin), Microsoft Bayesian Network Editor Format (.dsc), the Microsoft Bayesian Interchange Format (.bif) and XML BN Interchange Format (used by Netica, Hugin, and GeNie), and Ergo.


• Reimplementations of inference algorithms for Bayesian networks: clustering, variable elimination, conditioning (in progress); stochastic sampling


• Implementations of published algorithms and new algorithms for Bayesian network structure learning: K2 (Cooper and Herskovits, 1992), sparse candidate (Friedman et al., 1999), the genetic algorithm wrapper for K2 (GAWK - Hsu, Guo, Joehanes, Perry, Thornton, 2002)

We are now developing:

• New representations: relational (probabilistic relational models or PRMs), decision-theoretic (decision networks or influence diagrams), temporal (dynamic Bayesian networks or DBNs; hidden Markov models or HMMs), hybrid continuous state, continuous time


• New algorithms: PRM structure learning, decision network inference, factored frontier algorithms for DBNs, extensions of exact, sampling-based inference to continuous state and time


• Bioinformatics applications: Interfaces to software for computational genomics and other tools for computational biology

BNJ is available for download from SourceForge.

If you are interested in Java-based research and development tools for theoretical computer science and artificial intelligence, please feel free to join bayesnets for discussions and announcements.

Cheers,
William Hsu