From: tig@chello.pl   
      
   Hello.   
      
   I've got a problem.   
      
   During a test of back propagation algorithm I wanted to teach   
   the net the sin(x*y) function with a domain of [-2;2] for   
   both variables.   
      
   But it turned out that it can't be done.   
   The net manages to learn the function when the domain   
   is [-2;0] or [0;2] but it can't cope with [-2;2].   
      
   Learning with the domain of [-2;2] leads to zero at every output   
   and the better the net is learned the closer to zero every output   
   is.   
      
   My interpretation of that phenomenon is that the net   
   can't cope with the oddity of the sin function (but it can   
   do great with odd function of 1 variable e.g. sin(x)) and   
   because of that, during the learning, changes of weights   
   compensate themself leading to zero at the outputs.   
      
   What can be done to avoid it?   
      
   Regards.   
   TiGeR   
      
   [ comp.ai is moderated ... your article may take a while to appear. ]   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)