DEAL a software package for learning in Bayesian Networks

1
DEAL - a software package for learning in
Bayesian Networks

• Susanne G. BøttcherClaus Dethlefsen
• Aalborg University
• and
• Novo Nordisk A/S

2
The DEAL Team

• Susanne G. Bøttcher (alma_at_math.auc.dk)Soon to
  finish Ph.D. Supervisor Steffen Lauritzen
• Claus Dethlefsen (dethlef_at_math.auc.dk)Recent
  Ph.D. within state space models. Supervisor
  Søren Lundbye-Christensen
• Both supported by Novo Nordisk A/SAugust
  2001-August 2003

3
History of DEAL

• 1997-2002 Susannes Ph.D. Development of theory
• 2000-2001Birth Preliminary software in S-Plus
  (Susanne)
• Nov/Dec 2001 Prototype (Susanne/Claus)
• May 2002 Baptized
• Aug 2002 DEAL is released (Susanne/Claus)

4
What can DEAL do for us?
5
Bayesian Networks for Mixed Variables
6
Conditional Gaussian Distribution
7
Parameter and Structure Learning
8
Development

• Latent variables, missing observations
• Improve interface to Hugin
• Improvement on search strategies
• Graphical interface
• Applications in Diabetes research

9
Download

• www.math.auc.dk/novo/deal

10
R

• S-plus clone
• Free software (GNU)
• Open Source
• Runs on many platforms
• Supports dynamic loading of C-code
• http//sunsite.dk/R mirror of CRAN

11
gR initiative

• Steffen Lauritzen, Kurt Hornik, Peter Dalgaard
• Integrate facilities for computing with graphical
  models in R
• Open source and easily extendible

12
gR Software

• CoCo
• Digram
• Mim
• Tetrad
• Bugs
• Deal
• ...

13
Prediction of Insulin Sensitivity Index
14
Graphical model of Novos approach
15
It seems to be more complex
16
Demonstration