ControllerNet Class Reference

multi layer neural network with configurable activation functions and propagation and projection methods suitable for homeokinesis controller More...

#include <controllernet.h>

Inherits Configurable.

Collaboration diagram for ControllerNet:
Collaboration graph
[legend]

List of all members.

Public Member Functions

 ControllerNet (const std::vector< Layer > &layers, bool useBypass=false)
virtual ~ControllerNet ()
virtual void init (unsigned int inputDim, unsigned int outputDim, double unit_map=0.0, double rand=0.2, RandGen *randGen=0)
 initialisation of the network with the given number of input and output units.
virtual const matrix::Matrix process (const matrix::Matrix &input)
 passive processing of the input.
virtual const matrix::Matrix processX (const matrix::Matrix &input, const matrix::Matrix &injection, unsigned int injectInLayer)
 like process just with the opportunity to overwrite the activation of a specific layer
virtual void damp (double damping)
 damps the weights and the biases by multiplying (1-damping)
virtual const matrix::Matrix & response () const
 response matrix of neural network (for current activation, see process)

\[ J_ij = \frac{\partial y_i}{\partial x_j} \]

\[ J = G_n' W_n G_{n-1}' W_{n-1} ... G_1' W_1 \]

with $W_n$ is the weight matrix of layer n and $ G'$ is a diagonal matrix with $ G'_ii = g'_i $ as values on the diagonal.

virtual matrix::Matrix responsePart (int from, int to) const
 like response, just that only a range of layers is considered The Bypass is not considered here.
virtual const matrix::Matrix & responseLinear () const
 linear response matrix of neural network

\[ R = W_n W_{n-1} ... W_1 \]

with $W_n$ is the weight matrix of layer n.

virtual const matrix::Matrix backpropagation (const matrix::Matrix &error, matrix::Matrices *errors=0, matrix::Matrices *zetas=0) const
 backpropagation of vector error through network.
virtual const matrix::Matrix backpropagationX (const matrix::Matrix &error, matrix::Matrices *errors=0, matrix::Matrices *zetas=0, int startWithLayer=-1) const
 like backpropagation but with special features: we can start from any layer and the bypass-discounting can be used (see disseration Georg Martius) WARNING: the errors and zetas above the `startWithLayer' are undefined
virtual const matrix::Matrix backprojection (const matrix::Matrix &error, matrix::Matrices *errors=0, matrix::Matrices *zetas=0) const
 backprojection of vector error through network.
virtual const matrix::Matrix forwardpropagation (const matrix::Matrix &error, matrix::Matrices *errors=0, matrix::Matrices *zetas=0) const
 forwardpropagation of vector error through network.
virtual const matrix::Matrix forwardprojection (const matrix::Matrix &error, matrix::Matrices *errors=0, matrix::Matrices *zetas=0) const
 forwardprojection of vector error through network.
virtual unsigned int getInputDim () const
 returns the number of input neurons
virtual unsigned int getOutputDim () const
 returns the number of output neurons
virtual const matrix::Matrix & getLayerOutput (int layer) const
 returns activation of the given layer.
virtual unsigned int getLayerNum () const
virtual const Layer & getLayer (unsigned int layer) const
 layers 0 is the first hidden layer
virtual Layer & getLayer (unsigned int layer)
 layers 0 is the first hidden layer
virtual const matrix::Matrix & getWeights (int to_layer) const
 weight matrix 0 connects input with the first hidden layer Negative values count from the end (-1 is the last layer)
virtual matrix::Matrix & getWeights (int to_layer)
 weight matrix 0 connects input with the first hidden layer Negative values count from the end (-1 is the last layer)
virtual const matrix::Matrix & getByPass () const
virtual matrix::Matrix & getByPass ()
virtual const matrix::Matrix & getBias (int of_layer) const
 Note: layers 0 is the first hidden layer Negative values count from the end (-1 is the last layer).
virtual matrix::Matrix & getBias (int of_layer)
 Note: layers 0 is the first hidden layer Negative values count from the end (-1 is the last layer).
bool store (FILE *f) const
 stores the layer binary into file stream
bool restore (FILE *f)
 restores the layer binary from file stream
bool write (FILE *f) const
 writes the layer ASCII into file stream (not in the storable interface)

Protected Member Functions

virtual void calcResponseIntern ()

Protected Attributes

std::vector< Layer > layers
std::vector< matrix::Matrix > weights
std::vector< matrix::Matrix > bias
bool useBypass
matrix::Matrix bypassWeights
matrix::Matrix input
matrix::Matrices y
matrix::Matrices z
matrix::Matrices gp
matrix::Matrix L
matrix::Matrix R
double lambda
bool initialised

Detailed Description

multi layer neural network with configurable activation functions and propagation and projection methods suitable for homeokinesis controller

Constructor & Destructor Documentation

ControllerNet ( const std::vector< Layer > &  layers,
bool  useBypass = false 
)
Parameters:
layers Layer description (the input layer is not specified (always linear))
useBypass if true, then a connection from input to output layer is included
virtual ~ControllerNet (  )  [inline, virtual]

Member Function Documentation

virtual const matrix::Matrix backprojection ( const matrix::Matrix &  error,
matrix::Matrices *  errors = 0,
matrix::Matrices *  zetas = 0 
) const [virtual]

backprojection of vector error through network.

The storage for the intermediate values (errors, zetas) do not need to be given. The errors(layerwise) are at the output of the neurons (index 0 is at the input level, output of layer 0 has index 1 and so on) The zetas(layerwise) are the values inside the neurons that arise when backprojecting the error signal. (zeta[0] is at layer 0)

Returns:
errors[0] (result of backprojecting)
virtual const matrix::Matrix backpropagation ( const matrix::Matrix &  error,
matrix::Matrices *  errors = 0,
matrix::Matrices *  zetas = 0 
) const [virtual]

backpropagation of vector error through network.

The storage for the intermediate values (errors, zetas) do not need to be given. The errors(layerwise) are at the output of the neurons (index 0 is at the input level, output of layer 0 has index 1 and so on) The zetas(layerwise) are the values inside the neurons that arise when backpropagating the error signal. (zeta[0] is at layer 0)

Returns:
errors[0] (result of backpropagation)
virtual const matrix::Matrix backpropagationX ( const matrix::Matrix &  error,
matrix::Matrices *  errors = 0,
matrix::Matrices *  zetas = 0,
int  startWithLayer = -1 
) const [virtual]

like backpropagation but with special features: we can start from any layer and the bypass-discounting can be used (see disseration Georg Martius) WARNING: the errors and zetas above the `startWithLayer' are undefined

Parameters:
startWithLayer the error is clamped at this layer and the processing starts there (-1: output layer)
See also:
backpropagation
void calcResponseIntern (  )  [protected, virtual]
void damp ( double  damping  )  [virtual]

damps the weights and the biases by multiplying (1-damping)

virtual const matrix::Matrix forwardprojection ( const matrix::Matrix &  error,
matrix::Matrices *  errors = 0,
matrix::Matrices *  zetas = 0 
) const [virtual]

forwardprojection of vector error through network.

The storage for the intermediate values (errors, zetas) do not need to be given. The errors(layerwise) are at the output of the neurons (index 0 is at the input level = error, output of layer 0 has index 1 and so on) The zetas(layerwise) are the values inside the neurons that arise when forwardprojecting the error signal. (zeta[0] is at layer 0)

Returns:
errors[layernum] (result of forwardprojection)
virtual const matrix::Matrix forwardpropagation ( const matrix::Matrix &  error,
matrix::Matrices *  errors = 0,
matrix::Matrices *  zetas = 0 
) const [virtual]

forwardpropagation of vector error through network.

The storage for the intermediate values (errors, zetas) do not need to be given. The errors(layerwise) are at the output of the neurons (index 0 is at the input level = error, output of layer 0 has index 1 and so on) The zetas(layerwise) are the values inside the neurons that arise when forwardpropagate the error signal. (zeta[0] is at layer 0)

Returns:
errors[layernum] (result of forwardpropagation)
virtual matrix::Matrix& getBias ( int  of_layer  )  [inline, virtual]

Note: layers 0 is the first hidden layer Negative values count from the end (-1 is the last layer).

virtual const matrix::Matrix& getBias ( int  of_layer  )  const [inline, virtual]

Note: layers 0 is the first hidden layer Negative values count from the end (-1 is the last layer).

virtual matrix::Matrix& getByPass (  )  [inline, virtual]
virtual const matrix::Matrix& getByPass (  )  const [inline, virtual]
virtual unsigned int getInputDim (  )  const [inline, virtual]

returns the number of input neurons

virtual Layer& getLayer ( unsigned int  layer  )  [inline, virtual]

layers 0 is the first hidden layer

virtual const Layer& getLayer ( unsigned int  layer  )  const [inline, virtual]

layers 0 is the first hidden layer

virtual unsigned int getLayerNum (  )  const [inline, virtual]
virtual const matrix::Matrix& getLayerOutput ( int  layer  )  const [inline, virtual]

returns activation of the given layer.

Layer 0 is the first hidden layer. Negative values count from the end (-1 is the last layer)

virtual unsigned int getOutputDim (  )  const [inline, virtual]

returns the number of output neurons

virtual matrix::Matrix& getWeights ( int  to_layer  )  [inline, virtual]

weight matrix 0 connects input with the first hidden layer Negative values count from the end (-1 is the last layer)

virtual const matrix::Matrix& getWeights ( int  to_layer  )  const [inline, virtual]

weight matrix 0 connects input with the first hidden layer Negative values count from the end (-1 is the last layer)

void init ( unsigned int  inputDim,
unsigned int  outputDim,
double  unit_map = 0.0,
double  rand = 0.2,
RandGen *  randGen = 0 
) [virtual]

initialisation of the network with the given number of input and output units.

The dimensionality of the ouputlayer is automatically adjusted.

Parameters:
unit_map defines the approximate response of the network after initialisation (if unit_map=1 the weights are unit matrices).
randGen pointer to random generator, if 0 an new one is used
const Matrix process ( const matrix::Matrix &  input  )  [virtual]

passive processing of the input.

This has to be done before calling reponse, and the back/forward propagation/projection functions. The activations and the response matrix are stored internally.

const Matrix processX ( const matrix::Matrix &  input,
const matrix::Matrix &  injection,
unsigned int  injectInLayer 
) [virtual]

like process just with the opportunity to overwrite the activation of a specific layer

Parameters:
injections the input that is clamped at layer injectInLayer
injectInLayer the injection is clamped at this layer
const Matrix & response (  )  const [virtual]

response matrix of neural network (for current activation, see process)

\[ J_ij = \frac{\partial y_i}{\partial x_j} \]

\[ J = G_n' W_n G_{n-1}' W_{n-1} ... G_1' W_1 \]

with $W_n$ is the weight matrix of layer n and $ G'$ is a diagonal matrix with $ G'_ii = g'_i $ as values on the diagonal.

const Matrix & responseLinear (  )  const [virtual]

linear response matrix of neural network

\[ R = W_n W_{n-1} ... W_1 \]

with $W_n$ is the weight matrix of layer n.

Matrix responsePart ( int  from,
int  to 
) const [virtual]

like response, just that only a range of layers is considered The Bypass is not considered here.

Parameters:
from index of layer to start: -1 at input, 0 first hidden layer ...
to index of layer to stop: -1: last layer, 0 first hidden layer ...
bool restore ( FILE *  f  ) 

restores the layer binary from file stream

bool store ( FILE *  f  )  const

stores the layer binary into file stream

bool write ( FILE *  f  )  const

writes the layer ASCII into file stream (not in the storable interface)

Member Data Documentation

std::vector<matrix::Matrix> bias [protected]
matrix::Matrix bypassWeights [protected]
matrix::Matrices gp [protected]
bool initialised [protected]
matrix::Matrix input [protected]
matrix::Matrix L [protected]
double lambda [protected]
std::vector<Layer> layers [protected]
matrix::Matrix R [protected]
bool useBypass [protected]
std::vector<matrix::Matrix> weights [protected]
matrix::Matrices y [protected]
matrix::Matrices z [protected]
The documentation for this class was generated from the following files:
Generated on Thu Jun 28 14:48:08 2012 for Robot Simulator of the Robotics Group for Self-Organization of Control by  doxygen 1.6.3