Main Page

ANN - Artificial Neural Network for PHP 5.x

This project realizes a neural network topology called multilayer perceptron for PHP 5.x environments. The source code is based on a work by Eddy Young in 2002. Several improvements and changes on this implementation are done by Thomas Wien (thwien.de - Düsseldorf - Germany) since 2007. You will find the PHP source in the section Download. Please, consider the Copyright. To get a short idea what is the benefit of neural networks have a look at page Neural Networks.

Donate

If you want to support the current and future development of this project I would appreciate if you donate a freely amount via paypal.

Donate now

Overview

• Neural Networks
• Multilayer Perceptron
• Download
• Installation
• Examples
• Development
• FAQ
• Copyright

Features

• Output type detection (linear or binary)
• Logging (weights and network errors)
• Client-Server model for distributed applications
• Graphical network topology as PNG image
• Displaying network details
• String association
• Classification
• Phar support (as of PHP 5.3.0)

Versions and Change-Log

Version 2.1.x by Thomas Wien (Development)

• Image matrix support

Version 2.1.4 by Thomas Wien (2011-05-23) Download

* Better calculation of remaining time of running the network
* Fixing bug: generating random delta not correct
* Adding momentum value
* Simplified ANN_Neuron::adjustWeights()
* Remove possible wrong calculation in ANN_Neuron::adjustWeights() on linear networks
* Simplify ANN_Layer::calculateHiddenDeltas()

Version 2.1.3 by Thomas Wien (2010-01-06)

• Date input support class

Version 2.1.2 by Thomas Wien (2009-12-26)

• Classification support
• Phar support (as of PHP 5.3.0)

Version 2.1.1 by Thomas Wien (2009-12-23)

• String association support

Version 2.1.0 by Thomas Wien (2009-12-22)

• Checking parameter counts on ANN_Values::input() and ANN_Values::output()
• Removing protected method ANN_Neuron::getInputs()
• Fixing bug: Error tolerance calculation in ANN_Network::isTrainingComplete()
• Switching to Git version control
• Moving all experimental code into branch
• Removing all experimental code from master branch (due to performance and future development)

Version 2.0.7 by Thomas Wien (2009-01-01) Download

• Removing protected method ANN_Neuron::setOutput()
• Removing protected unused method ANN_Layer::getInputs()
• Removing protected unused property ANN_Layer::$arrInputs
• More detailed exceptions to ANN_Filesystem::saveToFile()
• Different distribution of activation calls across the layers
• Different adjustments in ANN_Neuron::adjustWeights() depending on output type
• Removing static local variables from ANN_Network::getNextIndexInputsToTrain()
• Increasing math precision
• Using class constants for output types (increasing performance)
• Fixing bug: ANN_Neuron::getOutput() is float and not array

Version 2.0.6 by Thomas Wien (2008-12-18)

• Printing network details of output differences to their desired values
• Complete rewritten code standard of variables
• New class ANN_Values for defining input and output values
• Code examples to phpdoc
• Internal math precision defaults to 5

Version 2.0.5 by Thomas Wien (2008-12-16)

• Adjustable output error tolerance between 0 and 10 per cent
• Internal rounding of floats for performance issues
• Loading class for all ANN classes (SPL autoload)
• Renaming filename of ANN_Maths class
• Improving code standard
• Fixing bug: Comparison in ANN_InputValue and ANN_OutputValue

Version 2.0.4 by Thomas Wien (2008-01-27)

• Weight decay
• QuickProp algorithm (experimental)
• RProp algorithm (experimental)
• Linear saturated activation function (experimental)
• Individual learning rate algorithm (experimental)
• Reducing of overfitting (no training if input pattern produces desired output)
• Increasing performance on activation
• Increasing performance on testing all patterns to their desired outputs
• Increasing performance on calculating hidden deltas
• Increasing performance by defining layer relation by construction
• More details to printNetwork()
• Fixing bug: learning rate is not part of saved delta value

Version 2.0.3 by Thomas Wien (2008-01-17)

• Support for dynamic learning rate
• Automatic epoch determination
• Automatic output type detection
• Shuffling input patterns each epoch instead of randomized pattern access
• Logging of network errors
• Logging on each epoch instead of each training step
• Avoiding distributed internal calls of setMomentum() and setLearningRate()
• Extending display of network details
• Fixing bug: runtime error on call of setMomentum()

Version 2.0.2 by Thomas Wien (2008-01-14)

• Client-Server model for distributed applications
• Calculating total network error for csv logging

Version 2.0.1 by Thomas Wien (2008-01-06)

• Separation of classes to several files
• Version control by Subversion
• Performance issues
• Graphical output of neural network topology
• Logging of weights to csv file

Version 2.0.0 by Thomas Wien (2007-12-17)

• PHP 5.x support
• PHPDoc documentation
• Momentum support
• Avoiding network overfitting
• Linear / binary output
• ANN_InputValue + ANN_OutputValue classes
• Exceptions
• Threshold function
• Hyperbolic tangent transfer function
• Several performance issues
• Avoiding array_keys() & srand() due to performance
• Changes in saving and loading network
• Printing network details to browser
• Fixing bug: initializing inputs to all hidden layers
• Fixing bug: training for first hidden layer was skipped

Version 1.0 by Eddy Young (2002)

• Initial version

Todo

• More Examples
• Performance check depending on host system
• Wiki: More details to installation and use
• PHPDoc: More details to documentation
• Supporting PHP 5.3 namespaces (later)
• Improving license agreement of source code
• Adding error codes to exceptions
• Exception if network error does not reach minimum