help for feature selection,pattern recognition

Marco e Stefano sor4011 at iperbole.bologna.it
Wed Jan 22 00:39:40 EST 1997


Hello I ' m Marc from Italy and I have a problem !
I ' m a student in electronical eng. at the University of Bologna , and 
I'm working
in a pattern recognition problem , in particular at feature selection.
I have data about flat foot and normal foot , measured by a dynamometric
piezoelectrical plates built by Kissler .
They record the three components of the ground reaction force 
during the normal step .The signal of Fx,Fy,Fz are made of 33 samples 
each.
I need only the best discriminant ones for pattern recognition ( with 
neural net) between flat and normal.
I'm using Matlab to try an exaustive research of the value of an index ( 
the divergence
in the sense of Marill and Green ) over a set of 33 instants.I have 400 
measures of 
each of them.From these 33 instants I have to select a subset in base at 
the divergence index,
 of a priori unknown size , and so I tried with an M-file of Matlab ALL 
the
 possible subsets of size 2 and 3 and 4 .
All the combination of 33 element taken at 4 at the time are a lot , about 
40.000 and for 5 element about 237.000.I wanted to try with subset of size 
from
5 to 10 , but just with 5 elements it takes too much time .
So I thinked to try with a language like C++ ( I use Borland C++ ver 1.00 
1990 ).
I built functions like mean of a matrix  , and cov of a matrix as in 
Matlab , but 
for the divergence I need also to calculate the inverse of the covariance 
matrices.
The source code of inv.m is not available and I don't know how to 
calculate the inverse
of a matrix , in C++ , in an efficient way .In fact the problem is 
calculating 
the inverse of covariance matrix : I ' ve 2 cov. matrices for each try so 
40.000 times 
I must get the inverse of each of them and this with subset of size 5 !
I cannot call the inv.m from C++ , and I want to try with subset of size 
>From 5 to 7
( with 8 there are 4.000.000 combinations ).
Could you help me or give just some suggestions ? 
Thank you for the attention , I will be very happy to have any type of 
answer.
Excuse me  for bad english .
Marco Mastrogiacomo   ( Email sor4011 at iperbole.bologna.it)

                                                Bologna Italy 17 January 
1997 
 





More information about the Comp-bio mailing list