changelog

递归贝叶斯估计的工具包

ReBEL ChangeLog
***************

This is the main ReBEL Changelog file documenting the ongoing evolution 
and debugging process of the ReBEL Toolkit. Below the changes from one
major or minor release to the next is outlined. The numbers in brackets
after each release version number is the number of subsequent 'bugfix' 
releases which has been made. So "Version 0.2 (1,2)" implies 2 bugfix 
versions has been released since the release of ReBEL-0.2, namely
ReBEL-0.2.1 and ReBEL-0.2.2. The details of each bugfix release is 
documented in more detail in the Bugfixes file which can be found the 
root ReBEL directory.

	
	
Version 0.2 (1,2,3,4,5,6)
===========

* Added new inference algorithms:

  - Gaussian Sum Particle Filter   (gspf)
  - Gaussian Mixture Sigma-Point Particle Filter (gmsppf) 
        <still in beta state>

  These hybrid particle filters try to improve on the generic bootstrap 
  particle filter. The GSPF (developed by J. Kotecha & P. Djuric) uses 
  a GMM representation of the posterior state distribution, efectively 
  smoothing out the particle based representation. This filter requires 
  the noise sources to be modeled as GMMs themselves. 

  The GMSPPF is a hybrid extension of the Sigma-Point Particle Filter 
  (sppf) that uses a SPKF-filterbank propagated GMM proposal distribution 
  to sample the particles from. The actual posterior state distribution 
  is also modeled with a GMM (like in the GSPF), but the difference is 
  that this GMM is fitted to the posterior weighted particle set (as 
  generated by the importance sampling step) by means of an EM algorithm.
  This eliminate the need for a variance increasing final resampling stage.
  This filter has equal (or better) performance compared to the SPPF, but 
  has a significantly lower computational cost. See code for more details.
	
* Added a new noise source type to the 'gennoiseds' function. This is
  a Gaussian mixture model noise source (type='gmm'). See gennoiseds.m
  for more detail.
	
* Added a new data structure (GMM) and functions to the core module of 
  ReBEL to support Gaussian mixture models (GMMs). These functions are:

    - gmmfit : Fit/train a GMM to data using the EM algorithm.
    - gmminitialize : Initiliaze a GMM (used internally by gmmfit)
    - gmmsample : Sample efficiently from a GMM
    - gmmprobability : calculate all probabilities of a GMM and a related 
	               dataset. 

  See the function definitions for more detail.
	
* The Netlab toolkit (neural network software for Matlab) is now bundled 
  with ReBEL. This is not a needed/crucial component of ReBEL, but 
  complements it nicely when certain neural network structures are needed
  within user defined models, etc. For more detail on Netlab, see
  http://www.ncrg.aston.ac.uk/netlab/ (Thanks to Ian Nabney and 
  Christopher Bishop for developing such a usefull resource). 

* Added more examples:

  - Dual Estimation : Added a speech enhancement demo based on dual SPKF 
    estimation. A speech fragment (phoneme), corrupted by additive
    colored Gaussian noise is cleaned up (filtered) through the use of a
    dual SPKF estimator. This example demonstrates how colored noise 
    is implemented withion the ReBEL framework as well as how dual 
    estimation is done.

* All particle filters : Changed the definition for all particle filters
  of the '.resampleThreshold' field in the InferenceDS data structure.
  It used to be an absolute number of particles, the threshold size
  of the effective particle set. This (the threshold) has now been  
  changed to a relative ratio of resampleThreshold = N_efective/N_total.
	
* Changes (actually additions) to the parameter estimation
  meta system blocks in 'geninfds.m':

  - Changed the paramFunSelect option of 'both' to 'both-p' which
    indicates the use of a parallel combination of FFUN and HFUN in
    the parameter estimation observation function, i.e.
          
          observ = |FFUN(X)|
                   |HFUN(X)|

  - Changed the meaning and implementation of the paramFunSelect
    option 'both'. This option (the default) now implies a serial
    concatenation of the original system state transition and state
    observation functions (FFUN and HFUN) to form the observation
    function for parameter estimation. The following is now implied
    for the paramFunSelect='both' option

          observ = HFUN(FFUN(X))

 
* Made changes to the NoiseDS data structure:

  - Changed '.ns_subtype' field to '.cov_type' (covariance type)
    Note : The '.subtype' field of the argument data structure used as
    input to the 'gennoiseds' function has also changed to '.cov_type'.

  - Changed covariance field name in all 'NoiseDS' data structures from
    '.P' and '.S' to '.cov'. This might seem ambiguous, but the type of
    covariance (full, sqrt, diag, etc.) is already clearly evident
    from the '.cov_type' field.

  - Removed the '.sqrt_flag' and '.diag_flag' fields. These where
    superfluous and did not serve the orginaly planned beneficial
    purpose.


* Bug fixes : Numerous bugs/typos have been fixed. Some of the more  
  serious ones were:

    - geninfds::linearize_generic : 
	 o Interface to function requires a varargout list which was 
	   not honored. This is now fixed.
	 o Internal indexing error.
    
    
Version 0.1 (1)
===========

* Initial Alpha release