view release on metacpan or  search on metacpan

lib/AI/ANN.pm  view on Meta::CPAN

easy for a user to specify the precise layout of their network (including both 
topology and weights, as well as many parameters), and to then retrieve those 
details. The purpose of this is to allow an additional module to then tweak 
these values by a means that models evolution by natural selection. The 
canonical way to do this is the included AI::ANN::Evolver, which allows 
the addition of random mutations to individual networks, and the crossing of 
two networks. You will also, depending on your application, need a fitness 
function of some sort, in order to determine which networks to allow to 
propagate. Here is an example of that system.

use AI::ANN;
my $network = new AI::ANN ( input_count => $inputcount, data => \@neuron_definition );
my $outputs = $network->execute( \@inputs ); # Basic network use
use AI::ANN::Evolver;
my $handofgod = new AI::ANN::Evolver (); # See that module for calling details
my $network2 = $handofgod->mutate($network); # Random mutations
# Test an entire 'generation' of networks, and let $network and $network2 be
# among those with the highest fitness function in the generation.
my $network3 = $handofgod->crossover($network, $network2);
# Perhaps mutate() each network either before or after the crossover to 
# introduce variety.

We elected to do this with a new module rather than by extending an existing 
module because of the extensive differences in the internal structure and the 
interface that were necessary to accomplish these goals. 

=head1 METHODS

=head2 new

ANN::new(input_count => $inputcount, data => [{ iamanoutput => 0, inputs => {$inputid => $weight, ...}, neurons => {$neuronid => $weight}}, ...])

input_count is number of inputs.
data is an arrayref of neuron definitions.
The first neuron with iamanoutput=1 is output 0. The second is output 1.
I hope you're seeing the pattern...
minvalue is the minimum value a neuron can pass. Default 0.
maxvalue is the maximum value a neuron can pass. Default 1.
afunc is a reference to the activation function. It should be simple and fast.
    The activation function is processed /after/ minvalue and maxvalue.
dafunc is the derivative of the activation function.
We strongly advise that you memoize your afunc and dafunc if they are at all
    complicated. We will do our best to behave.

=head2 execute

$network->execute( [$input0, $input1, ...] )

Runs the network for as many iterations as necessary to achieve a stable
network, then returns the output. 
We store the current state of the network in two places - once in the object,
for persistence, and once in $neurons, for simplicity. This might be wrong, 
but I couldn't think of a better way.

=head2 get_state

$network->get_state()

Returns three arrayrefs, [$input0, ...], [$neuron0, ...], [$output0, ...], 
corresponding to the data from the last call to execute().
Intended primarily to assist with debugging.

=head2 get_internals

$network->get_internals()

Returns the weights in a not-human-consumable format.

=head2 readable

$network->readable()

Returns a human-friendly and diffable description of the network.

=head2 backprop

$network->backprop(\@inputs, \@outputs)

Performs back-propagation learning on the neural network with the provided 
training data. Uses backprop_eta as a training rate and dafunc as the 
derivative of the activation function.

=head1 AUTHOR

Dan Collins <DCOLLINS@cpan.org>

=head1 COPYRIGHT AND LICENSE

This software is Copyright (c) 2011 by Dan Collins.

This is free software, licensed under:

  The GNU General Public License, Version 3, June 2007

=cut