NAME

AI::jNeural::nets::backprop - Backprop nets via libjneural

Synopsis

            

              
              
This documentation blows.  This module is about half done, so relax. ;)
Email feature requests, interface questions, angry emotive expressions, and 
all other commetns to the author.
Warning:  The interface can (and will) change at any time.  It will never do
less than it does now, but the calls may change drastically over time.  Deal
with it. ;)
            
          

new

            

              
              
$the_net = new AI::jNeural::nets::backprop($alpha, $inputs, $outputs, $hidden);
$alpha is the rate the weights are updated during back prop.
$inputs is the number of inputs
$outputs is the number of outputs
$hidden is the number of hidden layers.
   How many nodes per hidden layer?  
   K, this is dumb, but I'm working on it.
   x = (outputs + inputs);
   y = x+1;
   y = x-1;
   The number of nodes in each hidden layer are x,y,z,x,y,z,x,y,z ...
   libjneural takes a variable number of args to new(), until I work out 
   the XS vararg format... this is how it will work. ;)
            
          

train

            

              
              
This is a short cut method.  Well, it should be.  Presently it's the only
interface for the training and testing methods.  If you wanted to
personalize the training/testing then you'd need more methods.  Fred has
asked for this very feature.  We're working on it.
If I were you, I'd stick with the hashref interface (see below) since I'm
likely to retire the array arguments soon.
See below for what I hope is a realtively obvious calling syntax.
            
          

Example

            

              
              
use AI::jNeural::nets::backprop;
 my $alpha = 0.92;
 my ($inputs, $outputs) = (1, 1);
 my $hidden = 2;
 my $the_net;
 my $debug = 0;
 my $epsilon    = 0.0007;  # this defaults to 0.01
 my $max_epochs = 8000;  # this actually defaults to infinity
   
 $the_net = new AI::jNeural::nets::backprop($alpha, $inputs, $outputs, $hidden);
 # There is also BIPOLAR (among others iirc)
 $the_net->set_transfer(SIGMOID);
  
 # training data
 for my $n (0.1, 0.3, 0.5, 0.7, 0.9) {
     push @$sin_i1, [ $n ];
     push @$sin_o1, [ sin($n) ];
 }
 # test data
 for my $n (0.0, 1.12, 0.44, 0.85) {
     push @$sin_i2, [ $n ];
     push @$sin_o2, [ sin($n) ];
 }
 my ($min_error, $epochs_left) = 
     $the_net->train( $sin_i1, $sin_o1, $sin_i2, 
         $sin_o2, $epsilon, $max_epochs, $debug );
 # This may not be quite clear say there were 3 inputs and 2 outputs, then
 #
 # $in  = [ [1, 2, 3], [4, 5, 6], [7, 8, 9] ];
 # $out = [ [1, 2],    [3, 4],    [5, 6]    ];
 #
 # Follow?
 #
 # In any case, if you leave the test array_refs as undef's, then only the
 # training data will be used.
 # oh, and the above could have been:
 $the_net->train({
     training_inputs => $sin_i1,
      testing_inputs => $sin_i2,
     training_targets => $sin_o1,
      testing_targets => $sin_o2,
      epsilon=>$epsilon, alpha=>$alpha, max_epochs=>$max_epochs,
      debug=>1,
      training_loops_per_epoch => 10, 
          # defaults to 10
          # It's the number of times it trains on data before
          # it calculates an nmse and compares with the epsilon
 });
            
          

AUTHOR

Jettero Heller <jettero@voltar.org>

Jet's Neural Architecture is a C++ library. <http://www.voltar.org/jneural>

SEE ALSO

perl(1), AI::jNeural(3), AI::jNeural::arch::neuron(3), AI::jNeural::arch(3), AI::jNeural::nets::backprop(3), AI::jNeural::nets(3), AI::jNeural::nets::kohonen(3), AI::jNeural::nets::sarsa(3), AI::jNeural::utils::transfer(3), AI::jNeural::utils(3).