XPost: comp.ai.neural-nets, sci.image.processing, sci.math.num-analysis   
      
   Hi Milind,   
      
   Many thanks for your reply.  Sorry about my own tardy response.  I did   
   not immediately notice your reply.   
      
   > It does look like you've done some pretty good analysis. PCA might work   
   > well, and so would a whole list of other approaches. Have you tried   
   > boosted variants of simple algorithms?   
      
   What did you have in mind?  I did PCA by taking the correlation matrix   
   of the metrics and using R to get its eigenvectors and eigenvalues.   
   The eigenvalues had maxima around 2 and declined rather linearly down   
   to about 0.2 then fell more sharply.  I decided to use a kappa   
   (condition number) of 10 as a cutoff but I have read that the Kaiser   
   criterion, which uses an eigenvalue of 1 as a cutoff, is still widely   
   regarded as accepting too many components.   
      
   >   
   > Simple things go a long way: Ensure that you don't train on all your   
   > data, for example, and if possible, have some data that even you   
   > haven't seen before available.   
      
   I'm rather fortunate in that respect.  I can access a lot of new data.   
      
   >   
   > Another thing you could do is use some ready-made tools to do the basic   
   > tests, before you go on to your own implementation, this would save you   
   > a lot of time and speculation.   
      
   You are certainly correct.  I will try that when I get back from   
   Christmas break next week.   
      
   Thanks very much for your help,   
   Peter.   
      
   [ comp.ai is moderated.  To submit, just post and be patient, or if ]   
   [ that fails mail your article to , and ]   
   [ ask your news administrator to fix the problems with your system. ]   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)