view release on metacpan or  search on metacpan

lib/AI/Perceptron.pm  view on Meta::CPAN

              +---------------+
 X[1] o------ |W[1]      T    |
 X[2] o------ |W[2] +---------+         +-------------------+
  .           | .   |   ___   |_________|    __  Squarewave |_______\  Output
  .           | .   |   \     |    S    | __|    Generator  |       /
  .           | .   |   /__   |         +-------------------+
 X[n] o------ |W[n] |   Sum   |
              +-----+---------+

	     S  =  T + Sum( W[i]*X[i] )  as i goes from 1 -> n
	Output  =  1 if S > 0; else -1

Where C<X[n]> are the perceptron's I<inputs>, C<W[n]> are the I<Weights> that
get applied to the corresponding input, and C<T> is the I<Threshold>.

The I<squarewave generator> just turns the result into a positive or negative
number.

So in summary, when you feed the perceptron some numeric inputs you get either
a positive or negative output depending on the input's weights and a threshold.

=head1 TRAINING

Usually you have to train a perceptron before it will give you the outputs you
expect.  This is done by giving the perceptron a set of examples containing the
output you want for some given inputs:

    -1 => -1, -1
    -1 =>  1, -1
    -1 => -1,  1
     1 =>  1,  1

If you've ever studied boolean logic, you should recognize that as the truth
table for an C<AND> gate (ok so we're using -1 instead of the commonly used 0,
same thing really).

You I<train> the perceptron by iterating over the examples and adjusting the
I<weights> and I<threshold> by some value until the perceptron's output matches
the expected output of each example:

    while some examples are incorrectly classified
        update weights for each example that fails

The value each weight is adjusted by is calculated as follows:

    delta[i] = learning_rate * (expected_output - output) * input[i]

Which is know as a negative feedback loop - it uses the current output as an
input to determine what the next output will be.

Also, note that this means you can get stuck in an infinite loop.  It's not a
bad idea to set the maximum number of iterations to prevent that.

=head1 CONSTRUCTOR

=over 4

=item new( [%args] )

Creates a new perceptron with the following default properties:

    num_inputs    = 1
    learning_rate = 0.01
    threshold     = 0.0
    weights       = empty list

Ideally you should use the accessors to set the properties, but for backwards
compatability you can still use the following arguments:

    Inputs => $number_of_inputs  (positive int)
    N      => $learning_rate     (float)
    W      => [ @weights ]       (floats)

The number of elements in I<W> must be equal to the number of inputs plus one.
This is because older version of AI::Perceptron combined the threshold and the
weights a single list where W[0] was the threshold and W[1] was the first
weight.  Great idea, eh? :)  That's why it's I<DEPRECATED>.

=back

=head1 ACCESSORS

=over 4

=item num_inputs( [ $int ] )

Set/get the perceptron's number of inputs.

=item learning_rate( [ $float ] )

Set/get the perceptron's number of inputs.

=item weights( [ \@weights ] )

Set/get the perceptron's weights (floats).

For backwards compatability, returns a list containing the I<threshold> as the
first element in list context:

  ($threshold, @weights) = $p->weights;

This usage is I<DEPRECATED>.

=item threshold( [ $float ] )

Set/get the perceptron's number of inputs.

=item training_examples( [ \@examples ] )

Set/get the perceptron's list of training examples.  This should be a list of
arrayrefs of the form:

    [ $expected_result => @inputs ]

=item max_iterations( [ $int ] )

Set/get the perceptron's number of inputs, a negative value implies no maximum.

=back