---

PerceptronNetwork Class Reference

#include <PerceptronNetwork.h>

List of all members.

Public Member Functions
	PerceptronNetwork (vector< unsigned int > desc_layers, const char *network_name="unnamed", const ActivationFunction *fact=&ActivationFunctions::fact_tanh)
	~PerceptronNetwork (void)
	PerceptronNetwork (PerceptronNetwork &source)
void	setActivationFunction (PerceptronLayerType type, const ActivationFunction *fact)
void	save (fstream &fs) const
bool	load (fstream &fs)
void	randomizeParameters (const RandomFunction *weight_func, const RandomFunction *theta_func)
void	setInput (vector< double > &in)
void	setOptimalOutput (vector< double > &optimal)
vector< double >	getOutput (void) const
double	errorTerm (void) const
void	resetDiffs (void)
void	propagate (void)
void	backpropagate (void)
void	postprocess (void)
void	update (void)
void	dumpNetworkGraph (const char *filename) const
double	getLearningParameter (void) const
void	setLearningParameter (double epsilon)
double	getOptimalTolerance (void) const
void	setOptimalTolerance (double tolerance)
double	getWeightDecayParameter (void) const
void	setWeightDecayParameter (double factor)
double	getMomentumTermParameter (void) const
void	setMomentumTermParameter (double factor)
Public Attributes
const char *	name
vector< double >	input
vector< double >	output
Protected Attributes
vector< double >	output_optimal
vector< PerceptronLayer * >	layers
double	epsilon
double	opt_tolerance
double	weight_decay
double	momentum_term
Friends
class	WeightMatrix

---

Detailed Description

One multilayer perceptron network.

---

Constructor & Destructor Documentation

PerceptronNetwork::PerceptronNetwork (  vector< unsigned int >    desc_layers, const char *    network_name = "unnamed", const ActivationFunction *    fact = &ActivationFunctions::fact_tanh )

PerceptronNetwork constructor Construct a multilayer perceptron network by layer description given through desc_layers. The vector size specifies the number of layers, the individual elements the number of neurons within its layer. Parameters: desc_layers  Number of neurons within each layer, given the input layer as first and the output layer as last. Hence, the size of the vector must be at least two. network_name  Symbolic name of the network. Can be NULL. The name is pointer-copied, so we steal a pointer here. fact  Activation function to use for all neurons by default within the network. By default, its the tangens-hyperbolicus function. For individual layers, based one their type, the activation function can be changed using the changeActivation method.

PerceptronNetwork::~PerceptronNetwork (  void    )

PerceptronNetwork destructor Remove all layers stored within the network.

PerceptronNetwork::PerceptronNetwork (  PerceptronNetwork &    source )

Copy constructor. Parameters: source  Source object to be copied.

---

Member Function Documentation

void PerceptronNetwork::backpropagate (  void    )

Backpropagation algorithm for the entire network. Calculate all delta error signals in every layer and their neurons. Every neuron must have a proper input/output signal assigned, and the output_optimal training target result is used for calculation.

void PerceptronNetwork::dumpNetworkGraph (  const char *    filename )  const

Dump whole neural network as graph file in the GraphViz file format (www.graphviz.org). Parameters: filename  Name of the file to write the graph data to.

double PerceptronNetwork::errorTerm (  void    )  const

Calculate the errorterm for the network The error value for the current output is calculated. The current optimal output must be given prior to calling this function. The formula used is .

double PerceptronNetwork::getLearningParameter (  void    )  const

Getter for the epsilon parameter.

double PerceptronNetwork::getMomentumTermParameter (  void    )  const

Getter for the momentum term parameter.

double PerceptronNetwork::getOptimalTolerance (  void    )  const

Getter for the optimal tolerance parameter.

vector< double > PerceptronNetwork::getOutput (  void    )  const

Getter for the entire network output.

double PerceptronNetwork::getWeightDecayParameter (  void    )  const

Getter for the weight decay parameter.

bool PerceptronNetwork::load (  fstream &    fs )

Load the entire network from stream fs. Parameters: fs  Stream (input) to read the network from. Returns: True on success, false on failure.

void PerceptronNetwork::postprocess (  void    )

Postprocess algorithm for the entire network.

void PerceptronNetwork::propagate (  void    )

Propagation algorithm for the entire network. The input levels to the network have to be set using the setInput() method, before. Afterwards the output of the network can be obtained using the getOutput() method.

void PerceptronNetwork::randomizeParameters (  const RandomFunction *    weight_func, const RandomFunction *    theta_func )

Randomize the variable network parameters, weightings and theta values. Parameters: weight_func  Function to randomize weighting values. theta_func  Function to randomize theta parameters.

void PerceptronNetwork::resetDiffs (  void    )

Reset all learned parameters. Call after an update has been made.

void PerceptronNetwork::save (  fstream &    fs )  const

Save the entire network as text into the stream fs. Parameters: fs  Stream (output) to save the network to.

void PerceptronNetwork::setActivationFunction (  PerceptronLayerType    type, const ActivationFunction *    fact )

Set the activation function for layers, based on their type. Parameters: type  Type of layers that will use the new activation function. fact  Activation function to use.

void PerceptronNetwork::setInput (  vector< double > &    in )

Setter for the entire network input. Parameters: in  Vector of input signal levels. The size must equal the input layer neuron count.

void PerceptronNetwork::setLearningParameter (  double    epsilon )

Setter for the epsilon parameter. Parameters: epsilon  Learn parameter, in the range of 0.0 to 0.5.

void PerceptronNetwork::setMomentumTermParameter (  double    factor )

Setter for the momentum term parameter. Parameters: factor  Value between 0.5 and 0.9.

void PerceptronNetwork::setOptimalOutput (  vector< double > &    optimal )

Setter for the optimal requested network output. Must be used before any learning algorithm is called (backpropagate). Parameters: optimal  Optimal network output for the current input.

void PerceptronNetwork::setOptimalTolerance (  double    tolerance )

Setter for the optimal tolerance parameter. Parameters: tolerance  Value between 0.0 and 0.2.

void PerceptronNetwork::setWeightDecayParameter (  double    factor )

Setter for the weight decay parameter. Parameters: factor  Value in the range of 0.00005 to 0.0001.

void PerceptronNetwork::update (  void    )

Update algorithm for the entire network.

---

Friends And Related Function Documentation

friend class WeightMatrix [friend]

The GUI visualization class is declared friend to allow access to individual layers.

---

Member Data Documentation

double PerceptronNetwork::epsilon [protected]

Generic learn parameter epsilon, should be between 0.05 and 0.5.

vector<double> PerceptronNetwork::input

Input vector, fed into the first layer of the network. The number of elements must equal the number of neurons within the input layer.

vector<PerceptronLayer *> PerceptronNetwork::layers [protected]

Left-to-right list of layers within the network. Must be at least two (one input and one output layer), but can be arbitrary large. More than four layers does not archive any improvement of the network capabilities, though.

double PerceptronNetwork::momentum_term [protected]

Momentum term parameter, should be between 0.5 and 0.9.

const char* PerceptronNetwork::name

Symbolic name of the network. Never used by the methods themselves, but convenient to use.

double PerceptronNetwork::opt_tolerance [protected]

Optimal tolerance parameter, should be between 0.0 and 0.2.

vector<double> PerceptronNetwork::output

Output vector, resulting from propagation through the entire network. The number of elements equals the number of output neurons within the network.

vector<double> PerceptronNetwork::output_optimal [protected]

Optimal expected output vector, used for the learning process. For simple propagation it is not needed.

double PerceptronNetwork::weight_decay [protected]

Weight decay parameter, should be between 0.005 and 0.03.

---

The documentation for this class was generated from the following files:

• PerceptronNetwork.h
• PerceptronNetwork.cpp

---