Back-Propagation Algorithm implementation. More...

Inheritance diagram for BackPropagationAlgo:

Classes
class	cluster_deltas
	The struct of Clusters and Deltas.
Public Member Functions
	BackPropagationAlgo (NeuralNet *n_n, UpdatableList update_order, double l_r=0.1)
	Constructor.
	BackPropagationAlgo ()
	Default Constructor.
	~BackPropagationAlgo ()
	Destructor.
virtual double	calculateMSE (const Pattern &)
	Calculate the Mean Square Error respect to Pattern passed.
virtual void	configure (ConfigurationParameters &params, QString prefix)
	Configures the object using a ConfigurationParameters object.
void	disableMomentum ()
	Disable momentum.
void	enableMomentum ()
	Enable the momentum.
DoubleVector	getError (Cluster *)
	This method returns the deltas calculated by the Back-propagation Algorithm.
virtual void	learn (const Pattern &)
	Starts a single training step.
virtual void	learn ()
	a single step of learning algorithm
double	momentum () const
	return the momentum
double	rate () const
	return the learning rate
virtual void	save (ConfigurationParameters &params, QString prefix)
	Save the actual status of parameters into the ConfigurationParameters object passed.
void	setMomentum (double newmom)
	Set the momentum value.
void	setRate (double newrate)
	Set the learning rate.
void	setTeachingInput (Cluster *output, const DoubleVector &ti)
	Set the teaching input for Cluster passed.
void	setUpdateOrder (const UpdatableList &update_order)
	Set the order on which the error is backpropagated through the NeuralNet.
UpdatableList	updateOrder () const
	Return the order on which the error is backpropaget through the NeuralNet.
Static Public Member Functions
static void	describe (QString type)
	Add to Factory::typeDescriptions() the descriptions of all parameters and subgroups.
Protected Member Functions
virtual void	neuralNetChanged ()
	Configure internal data for backpropagation on NeuralNet setted.

Detailed Description

Back-Propagation Algorithm implementation.

Definition at line 41 of file backpropagationalgo.h.

Constructor & Destructor Documentation

BackPropagationAlgo	(	NeuralNet *	n_n,
		UpdatableList	update_order,
		double	l_r = `0.1`
	)

Constructor.

Parameters:

neural_network	the BaseNeuralNet neural network to train
update_order	the UpdatableList for the backpropagation sequence
learn_rate	the double learning rate factor

Definition at line 27 of file backpropagationalgo.cpp.

References BackPropagationAlgo::neuralNetChanged().

BackPropagationAlgo ( )

Default Constructor.

Definition at line 34 of file backpropagationalgo.cpp.

References BackPropagationAlgo::neuralNetChanged().

~BackPropagationAlgo ( )

Destructor.

Definition at line 41 of file backpropagationalgo.cpp.

Member Function Documentation

double calculateMSE ( const Pattern & pat ) [virtual]

Calculate the Mean Square Error respect to Pattern passed.

Implements LearningAlgorithm.

Definition at line 188 of file backpropagationalgo.cpp.

References NeuralNet::inputClusters(), Pattern::inputsOf(), farsa::mse(), LearningAlgorithm::neuralNet(), NeuralNet::outputClusters(), Pattern::outputsOf(), and NeuralNet::step().

void configure	(	ConfigurationParameters &	params,
		QString	prefix
	)		`[virtual]`

Configures the object using a ConfigurationParameters object.

From the file you can configure the parameters of the BackPropagation in this way:

 [aBackPropagationGroup]
 neuralnet = nameOfGroupOfTheNeuralNet
 rate = learningRate      ;if it's not present, default is 0.0 !!
 momentum = momentumRate  ;if it's not present, means momentum disabled
 order = cluster2 linker1 cluster1 ; order of Cluster and Linker on which the error is backpropagated

As you can note, there is no configuration parameters for loading the learning set from here. This is intended. You need to load separately the learning set and call the method learn on the loaded learning set.

You can do that creating groups like the following (see class Pattern):

 [learningSet:1]
 cluster:1 = cluster1
 inputs:1 = 1 2 3     ; input
 cluster:2 = cluster2
 outputs:2 = 2 4      ; desired output
 [learningSet:2]
 cluster:1 = cluster1
 inputs:1 = 2 4 6     ; input
 cluster:2 = cluster2
 outputs:2 = 4 8      ; desired output
 ...
 [learningSet:N]
 cluster:1 = cluster1
 inputs:1 = 10 20 30    ; input
 cluster:2 = cluster2
 outputs:2 = 20 40      ; desired output

And call the method LearningAlgorithm::loadPatternSet( params, "learningSet" )

Parameters:

params	the configuration parameters object with parameters to use
prefix	the prefix to use to access the object configuration parameters. This is guaranteed to end with the separator character when called by the factory, so you don't need to add one

Implements ParameterSettableWithConfigureFunction.

Definition at line 241 of file backpropagationalgo.cpp.

References ConfigurationParameters::getObjectFromGroup(), ConfigurationParameters::getObjectFromParameter(), ConfigurationParameters::getValue(), LearningAlgorithm::setNeuralNet(), ConfigurationParameters::startRememberingGroupObjectAssociations(), and ConfigurationParameters::stopRememberingGroupObjectAssociations().

void describe ( QString type ) [static]

Add to Factory::typeDescriptions() the descriptions of all parameters and subgroups.

Reimplemented from ParameterSettable.

Definition at line 280 of file backpropagationalgo.cpp.

References ParameterSettable::addTypeDescription(), ParameterSettable::RealDescriptor::def(), ParameterSettable::Descriptor::describeObject(), ParameterSettable::Descriptor::describeReal(), ParameterSettable::RealDescriptor::help(), ParameterSettable::ObjectDescriptor::help(), ParameterSettable::IsMandatory, ParameterSettable::RealDescriptor::limits(), ParameterSettable::ObjectDescriptor::props(), and ParameterSettable::ObjectDescriptor::type().

void disableMomentum ( ) [inline]

Disable momentum.

Definition at line 103 of file backpropagationalgo.h.

void enableMomentum ( )

Enable the momentum.

Definition at line 95 of file backpropagationalgo.cpp.

DoubleVector getError ( Cluster * cl )

This method returns the deltas calculated by the Back-propagation Algorithm.

These deltas are set every time new targets are defined for the output layer(s), which are then used to update network weights when the method learn() is called.
They are also useful to calculate the network performance, but for that it must be used outside the learning cycle (a full learning iteration, that corresponds to present the network with all the patterns of the train data set). For that you must call getError( Cluster * anyOutputCluster ) for each line of your training set (you'll get a DoubleVector with the deltas for each unit of the cluster considered).
Then you can use those values to calculate your desired performance measure.
For instance: if you use it to calculate the Mean Square Error (MSE) of the network for your train data set you must accumulate the square of the the getError( anyOutputCluster ) output for each line of the set, and at the end divide it by the length of that same set (by definition the MSE is the sum of the squared differences between the target and actual output of a sequence of values). Getting the Root Mean Squared Error (RMSE) from this is trivial (you just need to calculate the square root of the MSE).

Warning:: The data returned by getError( Cluster * ) is computed every time you set a new output target, which means every time you call the setTeachingInput( Cluster * anyOutputCluster, const DoubleVector & teaching_input ) method. If your network has more than one output layer you have to call setTeachingInput() for all the output clusters before calling getError() for any of the clusters.

Definition at line 86 of file backpropagationalgo.cpp.

void learn ( const Pattern & pat ) [virtual]

Starts a single training step.

Implements LearningAlgorithm.

Definition at line 172 of file backpropagationalgo.cpp.

References NeuralNet::inputClusters(), Pattern::inputsOf(), BackPropagationAlgo::learn(), LearningAlgorithm::neuralNet(), NeuralNet::outputClusters(), Pattern::outputsOf(), BackPropagationAlgo::setTeachingInput(), and NeuralNet::step().