Wish list for BNT
------------------

Rewriting Matlab BNT functions into C
-------------------------------------

- Loopy belief propagation
  belprop_inf_engine and pearl_inf_enigne should be sped up.
  As a test, reproduce the motion estimation results at
    www.merl.com/projects/scene_learning/index.html


Pure Matlab programming - extensions to BNT
----------------------------------------

- Add better support for undirected graphical models and factor
  graphs.

- Generalized linear models
  Integrate glmlab, so users can express generalized logistic
    regression, Possion models, etc. 
  http://www.sci.usq.edu.au/staff/dunn/glmlab/glmlab.html
  Pierpaolo Brutti has already done a lot of this - see the softmax class.

- Max a posteriori (MAP) learning
  Currently only tabular_CPD supports priors (Dirichlet).
  The gaussian_CPD class should support a Normal prior on the
  mean/weight matrix,   and a Wishart prior on the covariance.
  This would eliminate the need for ugly hacks like resetting
  singular covariance matrices. ML estimation is currently done in
     BNT/CPDs/@gaussian_CPD/maximize_params.m
  This could be initially simplified by assumign there is at most 1
  discrete parent. 

- Fully Bayesian modelling 
  Implement a way to let the user represent parameters as random
  variables (nodes). 
  These should have appropriate conjugate priors.
  The parameters will be parents to other CPDs.

- Robust conditional Gaussian code
  BNT currently implements the Lauritzen92 method for CG potentials,
  which is numerically 
  unstable and fairly slow. The Lauritzen99 method is much better, and
  should be implemented. (Shan.Huang@intel.com has made a first
  attempt, but it does not yet work.)
  
- N most probable assignments
  The Viterbi algorithm computes the most probable assignment to the
  hidden nodes of an HMM. 
  Nilsson98 generalized this to compute the top N most prob. assignments in
  an arbitrary BN, which should be implemented.

- Time series stuff
  Currently it is not possible to predict future observed nodes using
  the DBN code! In fact, online filtering, fixed lag smoothing and
  prediction are all poorly supported.
  This should be fixed.

- Variational algorithms
  Implement mean field and variational Bayes

- Plates
  Add a way to specify plates (Buntine94), which might be useful for
  automatically 
  partitioning training data, and is definitely useful for Bayesian methods.



IO stuff
--------

- BUGS interface
  Add a way to read/write BUGS files, and call BUGS from BNT.
     http://www.mrc-bsu.cam.ac.uk/bugs/ 
  This can be used to do inference on fully Bayesian models.

- File parser
  Ken Shan has written a Java program which converts a file stored in the 
  Bayes net interchange format (BIF) to a file containing a sequence
  of matlab/BNT 
  commands to construct an equivalent model
      http://www.digitas.harvard.edu/~ken/bif2bnt/
  It would nice to be able to call this (or some other) parser
  directly from matlab. 
  An ideal interface would be of the form
      bnet = load_bnet_from_file(filename)
  A similar function should be written for the XML-BIF format.
  The Bayes net repository discusses a variety of other popular file formats
     http://www.cs.huji.ac.il/labs/compbio/Repository/formats.html