/ 
AI-ParticleSwarmOptimization-MCE-1.005
River stage zero No dependents
/ AI::ParticleSwarmOptimization::MCE

NAME

AI::ParticleSwarmOptimization::MCE - Particle Swarm Optimization (object oriented) with support for multi-core processing

SYNOPSIS

use AI::ParticleSwarmOptimization::MCE;

my $pso = AI::ParticleSwarmOptimization::MCE->new (
    -fitFunc        => \&calcFit,
    -dimensions     => 3,
    -iterations     => 10,
    -numParticles   => 1000,
    
    # only for many-core version # the best if == $#cores of your system
    # selecting best value if undefined
    -workers		=> 4,							
);

my $fitValue       = $pso->optimize ();
my ($best)         = $pso->getBestParticles (1);
my ($fit, @values) = $pso->getParticleBestPos ($best);

printf "Fit %.4f at (%s)\n",
    $fit, join ', ', map {sprintf '%.4f', $_} @values;

sub calcFit {
    my @values = @_;
    my $offset = int (-@values / 2);
    my $sum;
    
    select( undef, undef, undef, 0.01 );    # Simulation of heavy processing...

    $sum += ($_ - $offset++) ** 2 for @values;
    return $sum;
}
=head1 Description

This module is enhancement of on original AI::ParticleSwarmOptimization to support multi-core processing with use of MCE. Below you can find original documentation of that module, but with one difference. There is new parameter "-workers", which one can use to define of number of parallel processes that will be used during computations.

The Particle Swarm Optimization technique uses communication of the current best position found between a number of particles moving over a hyper surface as a technique for locating the best location on the surface (where 'best' is the minimum of some fitness function). For a Wikipedia discussion of PSO see http://en.wikipedia.org/wiki/Particle_swarm_optimization.

This pure Perl module is an implementation of the Particle Swarm Optimization technique for finding minima of hyper surfaces. It presents an object oriented interface that facilitates easy configuration of the optimization parameters and (in principle) allows the creation of derived classes to reimplement all aspects of the optimization engine (a future version will describe the replaceable engine components).

This implementation allows communication of a local best point between a selected number of neighbours. It does not support a single global best position that is known to all particles in the swarm.

Methods

AI::ParticleSwarmOptimization provides the following public methods. The parameter lists shown for the methods denote optional parameters by showing them in [].

new (%parameters)

Create an optimization object. The following parameters may be used:

-workers: positive number, optional

The number of workers (processes), that will be used during computations.

-dimensions: positive number, required

The number of dimensions of the hypersurface being searched.

-exitFit: number, optional

If provided -exitFit allows early termination of optimize if the fitness value becomes equal or less than -exitFit.

-fitFunc: required

-fitFunc is a reference to the fitness function used by the search. If extra parameters need to be passed to the fitness function an array ref may be used with the code ref as the first array element and parameters to be passed into the fitness function as following elements. User provided parameters are passed as the first parameters to the fitness function when it is called:

my $pso = AI::ParticleSwarmOptimization::MCE->new(
    -fitFunc    => [\&calcFit, $context],
    -dimensions => 3,
);

...

sub calcFit {
    my ($context, @values) = @_;
    ...
    return $fitness;
}

In addition to any user provided parameters the list of values representing the current particle position in the hyperspace is passed in. There is one value per hyperspace dimension.

-inertia: positive or zero number, optional

Determines what proportion of the previous velocity is carried forward to the next iteration. Defaults to 0.9

See also -meWeight and -themWeight.

-iterations: number, optional

Number of optimization iterations to perform. Defaults to 1000.

-meWeight: number, optional

Coefficient determining the influence of the current local best position on the next iterations velocity. Defaults to 0.5.

See also -inertia and -themWeight.

-numNeighbors: positive number, optional

Number of local particles considered to be part of the neighbourhood of the current particle. Defaults to the square root of the total number of particles.

-numParticles: positive number, optional

Number of particles in the swarm. Defaults to 10 times the number of dimensions.

-posMax: number, optional

Maximum coordinate value for any dimension in the hyper space. Defaults to 100.

-posMin: number, optional

Minimum coordinate value for any dimension in the hyper space. Defaults to --posMax (if -posMax is negative -posMin should be set more negative).

-randSeed: number, optional

Seed for the random number generator. Useful if you want to rerun an optimization, perhaps for benchmarking or test purposes.

-randStartVelocity: boolean, optional

Set true to initialize particles with a random velocity. Otherwise particle velocity is set to 0 on initalization.

A range based on 1/100th of --posMax - -posMin is used for the initial speed in each dimension of the velocity vector if a random start velocity is used.

-stallSpeed: positive number, optional

Speed below which a particle is considered to be stalled and is repositioned to a new random location with a new initial speed.

By default -stallSpeed is undefined but particles with a speed of 0 will be repositioned.

-themWeight: number, optional

Coefficient determining the influence of the neighbourhod best position on the next iterations velocity. Defaults to 0.5.

See also -inertia and -meWeight.

-exitPlateau: boolean, optional

Set true to have the optimization check for plateaus (regions where the fit hasn't improved much for a while) during the search. The optimization ends when a suitable plateau is detected following the burn in period.

Defaults to undefined (option disabled).

-exitPlateauDP: number, optional

Specify the number of decimal places to compare between the current fitness function value and the mean of the previous -exitPlateauWindow values.

Defaults to 10.

-exitPlateauWindow: number, optional

Specify the size of the window used to calculate the mean for comparison to the current output of the fitness function. Correlates to the minimum size of a plateau needed to end the optimization.

Defaults to 10% of the number of iterations (-iterations).

-exitPlateauBurnin: number, optional

Determines how many iterations to run before checking for plateaus.

Defaults to 50% of the number of iterations (-iterations).

-verbose: flags, optional

If set to a non-zero value -verbose determines the level of diagnostic print reporting that is generated during optimization.

The following constants may be bitwise ored together to set logging options:

  • kLogBetter

    prints particle details when its fit becomes bebtter than its previous best.

  • kLogStall

    prints particle details when its velocity reaches 0 or falls below the stall threshold.

  • kLogIter

    Shows the current iteration number.

  • kLogDetail

    Shows additional details for some of the other logging options.

  • kLogIterDetail

    Shorthand for kLogIter | kLogIterDetail

setParams (%parameters)

Set or change optimization parameters. See -new above for a description of the parameters that may be supplied.

init ()

Reinitialize the optimization. init () will be called during the first call to optimize () if it hasn't already been called.

optimize ()

Runs the minimization optimization. Returns the fit value of the best fit found. The best possible fit is negative infinity.

optimize () may be called repeatedly to continue the fitting process. The fit processing on each subsequent call will continue from where the last call left off.

getParticleState ()

Returns the vector of position

getBestParticles ([$n])

Takes an optional count.

Returns a list containing the best $n particle numbers. If $n is not specified only the best particle number is returned.

getParticleBestPos ($particleNum)

Returns a list containing the best value of the fit and the vector of its point in hyper space.

my ($fit, @vector) = $pso->getParticleBestPos (3)
getIterationCount ()

Return the number of iterations performed. This may be useful when the -exitFit criteria has been met or where multiple calls to optimize have been made.

BUGS

None... I hope.

If any: A small script which yields the problem will probably be of help.

SEE ALSO

http://en.wikipedia.org/wiki/Particle_swarm_optimization

THANKS

Mario Roy for suggestions about efficiency.

AUTHOR

Strzelecki Lukasz <lukasz@strzeleccy.eu>

SEE ALSO

AI::ParticleSwarmOptimization AI::ParticleSwarmOptimization::Pmap

COPYRIGHT

Copyright (c) Strzelecki Lukasz. All rights reserved. This program is free software; you can redistribute it and/or modify it under the same terms as Perl itself.