Software for Learning Belief Networks from Data


I support the Blue Ribbon Online Free Speech Campaign!

[Manipulation | Learning from Data | Glossary | References ]


Software for Learning Belief Networks from Data

This document describes software for fitting graphical belief function models, and related modes such as Bayesian networks, influence diagrams, and probabilistic graphical models from data, or learning with existing model structures.

This page only describes software which fits models to data. For manipulating already constructed models, see the manipulating belief models page.

Each entry on this list contains information about features of the software and contact information. Pricing information includes the date the price was last updated as this information may vary from time to time (based on past performance, I'm rather slow about updating this list). Hyperlinks take you demo versions (or actual software for free software) when it is available on-line.

A key to terms used in the feature lists and a list of references are available below.


Packages reviewed in this document:

"+" indicates new since last version "*" indicates changed since last version This is version 1.4.


Scope, Purpose and Remarks

Building this list of software is more difficult than the corresponding one for fitting models because of the difficulty of defining the scope of this project. Clearly both general purpose statistical or knowledge discovery tools could fall under the umpbrella of this project. In particular, many log-linear models, path analysis and regression models are graphical models, so any statistics package capable of performing regression could potentially be included (Whittaker[1990] provides a good description of the link between statistics and graphical models). This is obviously far more work than I'm willing to entertain. Similarly, I've decided that packages which support tree-based models (e.g., CART) or neural networks are outside the scope of this page (although there may be enough conceptual similarity to justify their inclusion).

The packages described here specifically support graphical models with nodes representing variables and edges interactions in either their inputs or their outputs. People interested in other kinds of software are encouraged to check out these sources:

Also, a number of the packages for manipulating graphical belief models support updating of parameters from data. These include HUGIN and XBaies.

Free Software

Software in this first list is free for educational and research purposes. In some cases it may be more freely used. Each listing includes any restrictions on the use of the software.

Hypertext links to the actual software or product descriptions are provided where available.

BIFROST

Description:
BIFROST fits block recursive models (chain graphs) to data. It works with the program CoCo to do the fitting and will export models in a form that can be read by Hugin for analysis.
Scope:
Discrete Graphical Models, Chain Graphs
Features:
Model Selection, Export to Hugin
Platforms:
Runs on top of CoCo.
Documentation:
R92-2001, R92-2010
Program:
Bifrost1.1
Contact:
Bo Thiesson (thiesson@iesd.auc.dk)
[Return to Table Of Contents]

BNG

Description:
BNG isn't really software for learning belief networks, rather it is software for building belief networks
Scope:
Discrete Graphical Models
Features:
Knowledge Based Model Construction
Platforms:
Common Lisp, generates rule bases for IDEAL.
Program:
Software Descriptions
References:
Ngo, Haddawy and Helwig [1995], Haddawy et al[1995], Haddawy [1994].
Contact:
Peter Haddawy (haddawy@sprecher.cs.uwm.edu )
[Return to Table Of Contents]

BUGS

Description:
BUGS is a general purpose Markov Chain Monte Carlo program (Gibbs sampler) which because of its input syntax is particularly easy to use with graphical models.
Scope:
Graphical Models, Mixed Models, Latent Variables, Missing Data, General Bayesian Inference.
Features:
Comes with CODA a suit of S-Plus tools for assessing convergence.
Platforms:
Unix and Dos.
Restrictions:
Registration Required
Home Page:
England US Mirror
Documentation:
Documentation and Worked Examples.
Program:
Software Download and Registration
Contact:
bugs@mrc-bsu.cam.ac.ukat the MRC Biostatistics Unit
Notes:
If you have trouble accessing the files from here, try using ftp at ftp.mrc-bsu.cam.ac.uk and look in the directory /pub/metholodogy/bugs, or try the US Mirror
Warning:
Gibbs Sampling can be Dangerous! It allows you to fit any model you like to any data, no matter how stupid. Make sure you know what you're doing.
[Return to Table Of Contents]

CoCo

Description:
CoCo fits contingency tables to graphical models through a variety of selection procedures.
Scope:
Discrete Graphical Models
Features:
Model Selection, Graphical Interface in XLISP-STAT and link to S-Plus.
Platforms:
Macintosh (old version), Windows and Unix (latest)
Documentation:
See the CoCo Information Page
Program:
CoCo can be obtained from iesd.auc.dk in Denmark, or from the USA Mirror stat.ucla.edu.
Contact:
Jens Henrik Badsberg (jhb@iesd.auc.dk)
[Return to Table Of Contents]

COMMERCIAL

MIM

Description:
MIM is a general purpose package for fitting graphical and hierarchical models to mixed discrete and continuous data using maximum likelihood estimation.
Scope:
Graphical Models, Hierarchical Interaction Models.
Features:
GUI, Statistical Tests, Model Selection, Latent Variables, Diagnostic Graphics
Platforms:
DOS, Windows
Price:
Single-user License US $250, Multi-user License $1000, Student Version free. (Dec, 1995).
Demo:
student version (of MIM 2.3 for Windows)
Statlib Mirror
Contact:
David Edwards (DEd@novo.dk)
More information:
MIM Information Page
Reference:
Edwards [1995] (This book is the complete reference for MIM and contains a copy of the student version for DOS).
[Return to Table Of Contents]

TETRAD II

Description:
A multi-module program that assists in the construction of causal explanations for sample daata and their use in prediction.
Scope:
Discrete Bayesian Networks, Path Models and Structural Equation Models (directly estimates discrete networks, but creates input code to LISREL, EQS or CALIS, which are available elsewhere, for other models).
Features:
Model Search, Latent Variables, Simulation, Bayesian Updating
Platforms:
DOS
Price:
$150, DOS program with examples and user's maual
Contact:
Richard Scheines (Scheines@andrew.cmu.edu)

Or order from

Lawrence Erlbaum Associates
10 Industrial Ave
Mahwah, NJ 07430-22262
1-800-926-6579
1-201-236-9500
1-201-236-0072 (fax)
More Information:
The TETRAD Project Homepage
Reference:
Scheines \ea [1994] (This book is the reference for TETRAD and contains a copy of the software).
[Return to Table Of Contents]

[Manipulation | Learning from Data | Glossary | References ]


This list is maintained by Russell G. Almond.

Software for Fitting Belief Networks / almond@acm.org

Please send me any updates, corrections, omissions or suggestions for improvement. This is version 1.4 (3/2/95), so please bear with any mistakes. (As this is just a minor bug fix since the last release, I've left the "new" and "changed" markers from the 1.0 to 1.3 updagrade.) The next version should include information of model fitting software. Please send contributions to me.

Thanks to David Madigan and the University of Washington Statistics Department for hosting this Web page.


Last modified: Thu May 30 14:27:50 1996