==============================================================================
          FEATURE SELECTION IN STATISTICAL PATTERN RECOGNITION
==============================================================================
    This directory contains a collection of feature selection programs. 
    The idea underlying feature selection is to obtain features which 
maximize the similarity of objects in the same class while maximizing 
the dissimilarity of objects in different classes. This can be achieved 
by selecting features which maximize some distance measures.
    In these programs, two different distance measures are used:
1)  Mahalanobis distance (when the two covariance matrices are equal)
2)  Bhattacharyya distance (when the two covariance matrices are not equal)
If a program has "Jm" as a suffix, it simply indicates that the feature 
evaluation criterion chosen is the Mahalanobis distance measure (Jm). On 
the other hand, if a program has "Jb" as a suffix, it indicates that the 
Bhattacharyya distance (Jb) is chosen as the criterion function. Note that
these programs are designed for selecting features between "two" classes.
==============================================================================
                          Brief details of the programs
==============================================================================
exhaustiveJm, exhaustiveJb : programs performing exhaustive search to
                             obtain an optimal solution. It requires	
                             a lot of memory and computational time.
bbJm, bbJb                 : A method known as the "branch and bound (bb)" 
			     method can be used to obtain an optimal solution 
			     provided that the monotonicity condition 
			     is satisfied. It also requires a lot of 
			     memory and computational time but much
			     less than those of exhaustive search.
sfsJm, sfsJb               : Suboptimal solutions can be obtained using
			     sequential forward selection (sfs) method. In 
			     this method, one feature at a time is added to 
			     the currently selected feature set. It suffers 
			     from the nesting effect, i.e. once features are 
			     included into the currently selected feature
			     set it cannot be removed if it is later found 
			     to be superfluous. 
sbsJm, sbsJb               : A counterpart to SFS method is known as "the 
			     sequential backward selection (sbs) method".  
			     In this case, one feature at a time is deleted 
			     from the set of available features. It also 
			     suffers from the nesting effect, i.e. once features 
			     are removed, they cannot be added to the currently 
			     selected feature set again.
plmrJm, plmrJb             : This "plus l minus r (plmr) method" can prevent
			     nesting effect but it does not necessarily mean
			     that the selected features are optimal. It is 
			     basically a combination of SFS and SBS. If the
			     values of "l" is chosen to be greater than "r"
			     then it is a bottom up process. On the other hand,
			     if the values of "l" is chosen to be smaller than
			     "r" then it is a top down process. 
sffsJm, sffsJb             : This method is known as the "sequential forward
			     floating selection (sffs) method". This method 
			     can prevent the nesting effect without having to
			     decide the values of "l" and "r" as in the plmr
			     method.
sbfsJm, sbfsJb             : This is a counterpart method to the SFFS method.
                             It is known as the "sequential backward floating
			     selection (sbfs) method".
==============================================================================
INPUT FORMAT : 
               (a) when the covariance matrices are not equal, the format of
                   the input file is as follows:
                   Number of total features (Header)
                   Mean vector of class 1
                   Covariance matrix of class 1
                   Mean vector of class 2
                   Covariance matrix of class 2
                   
               (b) when the covariance matrices are equal, the format of the
                   input file is as follows:
                   Number of total features (Header)
                   Mean vector of class 1
                   Mean vector of class 2
                   Common covariance matrix 
 
==============================================================================
BUGS : When running the branch and bound, segmentation faults may occasionally 
       occur. These are due to the memory is not enough.
==============================================================================
For more details : see TR 1/91 or Nam Choakjarenwanit's 1992 PhD thesis. 
==============================================================================
<HR><ADDRESS>
06-May-97. Automatically converted by man2html,
 written by G.Matas (g.matas@ee.surrey.ac.uk)</ADDRESS>
</body>
</html>