GNG Class Reference

#include <GNG.h>

Inheritance diagram for GNG:

List of all members.

Public Member Functions
	GNG (classifierparams &params)
virtual	~GNG ()
virtual void	init (FeatureContainer *fc)
virtual membershiplist	getClusterMembership (const featurevector *sample)
virtual unsigned long	nextSample ()
virtual unsigned long	nextSample (const featurevector *sample)
virtual string	serialize () const
virtual void	unserialize (string data)
virtual double	getWinnerDistance (const featurevector *sample, unsigned int &cluster_id, unsigned int &meta_cluster_id) const
virtual unit_list::size_type	getNumberNodes () const
virtual meta_cluster_list	getMetaClusters ()
virtual string	toString () const
Static Public Attributes
static unsigned	lambda = 10
	insert a new unit after (lambda * <number of units>) samples
static double	minimal_deletion_age = 0.5
	only delete units which are older than that (their youth must be lower than this value)
static double	deletion_threshold = 0.5
	the threshold for the deletion criterion (it must be lower than this for a node to get deleted)
static double	deletion_threshold_lq = 0.15
	the threshold for the quality measure for learning when deleting a node (it must be lower than this for a node to get deleted)
static unsigned long	edge_age_max = 30
	maximum age of an edge
static double	T_short = 6
	time constant for exponential decrease of the short term error of a winning unit
static double	T_long = 170
	time constant for exponential decrease of the long term error of a winning unit
static double	T_youth = 50
	time constant for exponential decrease of the youth
static double	T_insert_threshold = 55
static double	adaptation_threshold = 0.0005
static double	insertion_tolerance = 0.02
static double	lr_winner = 0.15
	the learning rate of the winning unit
static double	lr_neighbor = 0.01
	the learning rate of all neighbors
static double	minimum_lr = 0.2
	the absolute minimum learning rate
static double	lr_insert_threshold = 0.2
	the learning rate for the insertion threshold
static double	helper_T_short = -1
static double	helper_T_long = -1
static double	helper_T_youth = -1
static double	helper_T_insertion_threshold = -1

---

Detailed Description

This is lifelong GNG. It is basically an implementation of Fred H. Hamker: "Life-long learning Cell Structures---continuously learning without catastrophic interference" Neural Networks 12 (2001) 551--573

Todo:

documentation

---

Constructor & Destructor Documentation

GNG::GNG (  classifierparams &  params  )

Only stores the parameters for later use in init(). Parameters: params  Initialization parameters See also: init

GNG::~GNG (   )  [virtual]

Destructor

---

Member Function Documentation

membershiplist GNG::getClusterMembership (  const featurevector *  sample  )  [virtual]

Get membership vector This method returns the data for the next steps, which is a list of context classes found by the algorithm with associated membership values. These values specify the membership of the current feature values to each of the context classes in the interval [0; 1]. They can thus by seen as the probabilities that the context classes are currently active. Parameters: sample  The sample vector for which the membershiplist should be calculated Implements ClassifierAlgorithm.

meta_cluster_list GNG::getMetaClusters (   )  [virtual]

Todo: THIS MUST WORK !!!!

unit_list::size_type GNG::getNumberNodes (   )  const [virtual]

Todo: documentation

double GNG::getWinnerDistance (  const featurevector *  sample, unsigned int &  cluster_id, unsigned int &  meta_cluster_id )  const [virtual]

Todo: documentation

void GNG::init (  FeatureContainer *  fc  )  [virtual]

This method initializes the helper variables, creates two initial cluster center prototypes by calling getFeatureVector on the given FeatureContainer and stored the FeatureContainer reference for calls to nextSample(). Additionally, it initializes the bit vector for masking feature dimensions by using the given algorithm parameters and the feature dimensions from the retrieved feature vectors. See also: nextSample() Implements ClassifierAlgorithm.

unsigned long GNG::nextSample (  const featurevector *  inSample  )  [virtual]

Todo: cope with increasing dimension of the feature vector ? Todo: this should be done in a better way Implements ClassifierAlgorithm.

unsigned long GNG::nextSample (   )  [virtual]

Fetch next sample vector Gets the next sample vector from the configured feature source. This method calls nextSample and getSampleVector on the FeatureContainer given to the init method. See also: FeatureConteiner::nextSample FeatureContainer::getSampleVector Implements ClassifierAlgorithm.

string GNG::serialize (   )  const [virtual]

Serialize a samples data to a string Returns: String representation of the samples data. Implements ClassifierAlgorithm.

string GNG::toString (   )  const [virtual]

This is only for testing. Implements ClassifierAlgorithm.

void GNG::unserialize (  string  data  )  [virtual]

Unserialize a samples data from a string Parameters: data  String representation of the samples data. Implements ClassifierAlgorithm.

---

Member Data Documentation

unsigned GNG::lambda = 10 [static]

insert a new unit after (lambda * <number of units>) samples GNG parameter Todo: documentation

---

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

• amw/src/cpp/cross/contextprediction/framework/Classifiers/gng/GNG.h
• amw/src/cpp/cross/contextprediction/framework/Classifiers/gng/GNG.cpp

---