#include <NeuralNetwork.h>

Inheritance diagram for NeuralNetwork:

Collaboration diagram for NeuralNetwork:

Classes
struct	Layer
	Layer structure. More...

Public Member Functions
	NeuralNetwork (unsigned int nInputs, unsigned int nHiddens, unsigned int nOutputs, float learningRate=0.01, float decreaseConstant=0, float weightDecay=0, bool linearOutput=false)

virtual	~NeuralNetwork ()

virtual void	init ()
	Initializes the network (resets the weights, among other things). More...

virtual unsigned int	nInputs () const
	Returns the number of inputs. More...

virtual unsigned int	nHidden () const
	Returns the number of hidden neurons. More...

virtual unsigned int	nOutputs () const
	Returns the number of outputs. More...

virtual unsigned int	nParams () const
	Returns the number of parameters. More...

virtual float	getCurrentLearningRate () const
	Returns the current effective learning rate (= learningRate / (1 + t * decreaseConstant). More...

virtual void	setInput (int i, real x)
	Sets input i to value x. More...

virtual void	setInputs (const real *inputs)
	Sets the value of the inputs. More...

virtual real	getOutput (int i) const
	Get output i. More...

virtual void	getOutputs (real *outputs) const
	Get the value of the outputs. More...

virtual void	backpropagate (real *outputError)
	Backpropagates the error, updating the derivatives. More...

virtual void	propagate ()
	Propagates inputs to outputs. More...

virtual void	update ()
	Updates the weights according to the derivatives. More...

virtual void	save (XFile *file)
	Saves the model to a file. More...

virtual void	load (XFile *file)
	Loads the model from a file. More...

void	_allocateLayer (Layer &layer, unsigned int nInputs, unsigned int nOutputs, unsigned int &k, bool isLinear=false)

void	_deallocateLayer (Layer &layer)

void	_propagateLayer (Layer &lower, Layer &upper)

void	_backpropagateLayer (Layer &upper, Layer &lower)

void	_deallocate ()

Public Member Functions inherited from GradientFunction
	GradientFunction ()

virtual	~GradientFunction ()

virtual void	clearDelta ()
	Clears the derivatives. More...

Public Member Functions inherited from Function
	Function ()

virtual	~Function ()

Public Attributes
float	learningRate

float	decreaseConstant

float	weightDecay

float	_learningRateDiv

unsigned int	_nParams
	Number of parameters. More...

Layer	inputLayer
	The three MLP layers (inputs -> hidden -> outputs). More...

Layer	hiddenLayer

Layer	outputLayer

Public Attributes inherited from GradientFunction
real *	weights
	The weights (parameters) of the gradient function. More...

real *	dWeights
	The derivatives of the weights. More...

Constructor & Destructor Documentation

NeuralNetwork::NeuralNetwork	(	unsigned int	nInputs,
		unsigned int	nHiddens,
		unsigned int	nOutputs,
		float	learningRate = `0.01`,
		float	decreaseConstant = `0`,
		float	weightDecay = `0`,
		bool	linearOutput = `false`
	)

NeuralNetwork::~NeuralNetwork ( )

virtual

Member Function Documentation

void NeuralNetwork::_allocateLayer	(	Layer &	layer,
		unsigned int	nInputs,
		unsigned int	nOutputs,
		unsigned int &	k,
		bool	isLinear = `false`
	)

void NeuralNetwork::_backpropagateLayer	(	Layer &	upper,
		Layer &	lower
	)

void NeuralNetwork::_deallocate ( )

void NeuralNetwork::_deallocateLayer ( Layer & layer )

void NeuralNetwork::_propagateLayer	(	Layer &	lower,
		Layer &	upper
	)

void NeuralNetwork::backpropagate ( real * outputError )

virtual

Backpropagates the error, updating the derivatives.

Implements GradientFunction.

float NeuralNetwork::getCurrentLearningRate ( ) const

virtual

Returns the current effective learning rate (= learningRate / (1 + t * decreaseConstant).

real NeuralNetwork::getOutput ( int i ) const

virtual

Get output i.

Implements Function.

void NeuralNetwork::getOutputs ( real * output ) const

virtual

Get the value of the outputs.

Reimplemented from Function.

void NeuralNetwork::init ( )

virtual

Initializes the network (resets the weights, among other things).

Reimplemented from Function.

void NeuralNetwork::load ( XFile * file )

virtual

Loads the model from a file.

Reimplemented from GradientFunction.

virtual unsigned int NeuralNetwork::nHidden ( ) const

inlinevirtual

Returns the number of hidden neurons.

virtual unsigned int NeuralNetwork::nInputs ( ) const

inlinevirtual

Returns the number of inputs.

Implements Function.

virtual unsigned int NeuralNetwork::nOutputs ( ) const

inlinevirtual

Returns the number of outputs.

Implements Function.

virtual unsigned int NeuralNetwork::nParams ( ) const

inlinevirtual

Returns the number of parameters.

Implements GradientFunction.

void NeuralNetwork::propagate ( )

virtual

Propagates inputs to outputs.

Implements Function.

void NeuralNetwork::save ( XFile * file )

virtual

Saves the model to a file.

Reimplemented from GradientFunction.

void NeuralNetwork::setInput	(	int	i,
		real	x
	)

virtual

Sets input i to value x.

Implements Function.

void NeuralNetwork::setInputs ( const real * input )

virtual

Sets the value of the inputs.

Reimplemented from Function.

void NeuralNetwork::update ( )

virtual

Updates the weights according to the derivatives.

Implements GradientFunction.

Member Data Documentation

float NeuralNetwork::_learningRateDiv

This value is used to keep track of the learning rate divider: it is equal to (1 + t * decreaseConstant). It is more efficient than the usual way of updating the learning rate, because it requires only one floating point addition per iteration, instead of an addition and a multiplication.

unsigned int NeuralNetwork::_nParams

Number of parameters.

float NeuralNetwork::decreaseConstant

The learning rate decrease constant. Value should be >= 0, usually in [0, 1]. The decrease constant is applied as a way to slowly decrease the learning rate during gradient descent to help convergence to a better minimum.

Layer NeuralNetwork::hiddenLayer

Layer NeuralNetwork::inputLayer

The three MLP layers (inputs -> hidden -> outputs).

float NeuralNetwork::learningRate

The starting learning rate. Value should be >= 0, usually in [0, 1]. The learning rate is used to adjust the speed of training. The higher the learning rate the faster the network is trained. However, the network has a better chance of being trained to a local minimum solution. A local minimum is a point at which the network stabilizes on a solution which is not the most optimal global solution. In the case of reinforcement learning, the learning rate determines to what extent the newly acquired information will override the old information. A factor of 0 will make the agent not learn anything, while a factor of 1 would make the agent consider only the most recent information. Source: http://pages.cs.wisc.edu/~bolo/shipyard/neural/tort.html http://en.wikipedia.org/wiki/Q-learning#Learning_rate

Layer NeuralNetwork::outputLayer

float NeuralNetwork::weightDecay

The weight decay. Value should be >= 0, usually in [0, 1]. Weight decay is a simple regularization method that limits the number of free parameters in the model so as to prevent over-fitting (in other words, to get a better generalization). In practice, it penalizes large weights and thus also limits the freedom in the model.

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

src/qualia/learning/NeuralNetwork.h
src/qualia/learning/NeuralNetwork.cpp

Classes

Public Member Functions

Public Attributes

Constructor & Destructor Documentation

Member Function Documentation

Member Data Documentation