用于研究贝叶斯网络和图形模型的软件包

Software Packages for Graphical Models / Bayesian Networks

Written by Kevin Murphy.
Last updated 28 July 2008.

Review articles

• Click here for a short article I wrote for the ISBA (International Society for Bayesian Analysis) Newsletter, December 2007
• Click here for a more detailed discussion of some of these packages written by Ann Nicholson and Kevin Korb in 2004.
• Click here for a French version of my comparison table (not necessarily up-to-date).

What do the headers in the table mean?

• Src = source code included? (N=no) If so, what language?
• API = application program interface included? (N means the program cannot be integrated into your code, i.e., it must be run as a standalone executable.)
• Exec = Executable runs on W = Windows (95/98/NT), U = Unix, M = Mac, or - = any machine with a compiler.
• Cts = are continuous (latent) nodes supported? G = (conditionally) Gaussians nodes supported analytically, Cs = continuous nodes supported by sampling, Cd = continuous nodes supported by discretization, Cx = continuous nodes supported by some unspecified method, D = only discrete nodes supported.
• GUI = Graphical User Interface included?
• Learns parameters?
• Learns structure? CI = means uses conditional independency tests
• Utility = utility and decision nodes (i.e., influence diagrams) supported?
• Free? 0 = free (although possibly only for academic use). $ = commercial software (although most have free versions which are restricted in various ways, e.g., the model size is limited, or models cannot be saved, or there is no API.)
• Undir? What kind of graphs are supported? U = only undirected graphs, D = only directed graphs, UD = both undirected and directed, CG = chain graphs (mixed directed/undirected).
• Inference = which inference algorithm is used? jtree = junction tree, varelim = variable (bucket) elimination, MH = Metropols Hastings, G = Gibbs sampling, IS = importance sampling, sampling = some other Monte Carlo method, polytree = Pearl's algorithm restricted to a graph with no cycles, VMP = variational message passing, EP = expectation propagation, SL = the program is designed for structure learning from completely observed data, not state estimation
• Comments. If in "quotes", I am quoting the authors at their request.

If you want your package to be listed, please send me email in the format specified here (use view source to see this properly).

Name

Authors

Src

API

Exec

Cts

GUI

Params

Struct

Utility

Free

Undir

Inference

Comments

AgenaRisk

Agena

N

Y

W,U

Cx

Y

Y

N

N

$

D

JTree

Simulation by Dynamic discretisation

Analytica

Lumina

N

Y

W,M

G

Y

N

N

Y

$

D

sampling

spread sheet compatible

Banjo

Hartemink

Java

Y

W,U,M

Cd

N

N

Y

N

0

D

none

structure learning of static or dynamic networks of discrete variables

Bassist

U. Helsinki

C++

Y

U

G

N

Y

N

N

0

D

MH

Generates C++ for MCMC. (No longer maintained)

BayesBuilder

Nijman (U. Nijmegen)

N

N

W

D

Y

N

N

N

0

D

?

-

BayesiaLab

Bayesia Ltd

N

N

-

Cd

Y

Y

Y

N

$

CG

jtree,G

Structural learning, adaptive questionnaires, dynamic models

Bayesware Discoverer

Bayesware

N

N

WUM

Cd

Y

Y

Y

N

$

D

?

Uses bound and collapse for learning with missing data.

B-course

U. Helsinki

N

N

WUM

Cd

Y

Y

Y

N

0

D

?

Runs on their server: view results using a web browser.

Belief net power constructor

Cheng (U.Alberta)

N

W

W

D

Y

Y

CI

N

0

D

?

-

BayesBlocks

Helsinki

Python/C++

Y

-

Y

N

Y

N

N

0

Dir

Variational

Non-Gaussian Latent variable models

Blaise

Bonnowitz and Mansinghka

Java

Y

-

Y

N

Y

N

N

0

Fgraph

MCMC, SMC

General MC toolkit, also handles non-parametric Bayesian models

BNT

Murphy (U.C.Berkeley)

Matlab/C

Y

WUM

G

N

Y

Y

Y

0

D,U

Many

Also handles dynamic models, like HMMs and Kalman filters.

BNJ

Hsu (Kansas)

Java

-

-

D

Y

N

Y

N

0

D

jtree, IS

-

BNL

frank rijmen

Matlab

-

-

D

N

N

N

N

0

D

jtree

Supports (ordinal) logistic regression CPDs and EM learning

BUGS

MRC/Imperial College

N

N

WU

Cs

W

Y

N

N

0

D

Gibbs

-

Causal discoverer

Vanderbilt

N

N

W

-

-

N

Y

N

0

D

-

structure learning only

CoCo+Xlisp

Badsberg (U. Aalborg)

C/lisp

Y

U

D

Y

Y

CI

N

0

U

Jtree

Designed for contingency tables.

CIspace

Poole et al. (UBC)

Java

N

WU

D

Y

N

N

N

0

D

Varelim

-

CRFtoolbox

Schmidt and Murphy

Matlab/C

Y

-

N

N

Y

N

N

0

U

Loopy BP

Conditional random fields, arbitrary structure

DBNbox

Roberts et al

Matlab

-

-

Y

N

Y

N

N

Y

D

Various

DBNs

Deal

Bottcher et al

R

-

-

G

Y

Y

Y

N

0

D

None

Structure learning.

DeriveIt

DeriveIt LLC

N

-

-

?

?

Y

Y

?

$

D

Jtree, Gibbs

Exploits local structure in CPDs.

Elvira

Elvira consortium (Spain)

Java

Y

W,U,M

Cd,Cx

Y

Y

Y

Y

0

D

JTree,varelim,IS

"Also includes classification, abductive inference and model fusion"

Ergo

Noetic systems

N

Y

W,M

D

Y

N

N

N

$