Description of the class TGPBuilder. More...

#include <TGPBuilder.h>

Collaboration diagram for URANIE::Modeler::TGPBuilder:

[legend]

Public Member Functions
Constructor and Destructor
	TGPBuilder ()
	Default Constructor.

	TGPBuilder (URANIE::DataServer::TDataServer tdsObs, const char varexpinput, const char varexpoutput, TCorrelationFunction corrFunc, const char trend="", Option_t option="")
	Construction from a data server, a list of input and output variables and a correlation function object.

	TGPBuilder (URANIE::DataServer::TDataServer tdsObs, const char varexpinput, const char varexpoutput, const char corrFunc, const char trend="", Option_t option="")
	Construction from a data server, a list of input and output variables and a correlation function name.

	TGPBuilder (URANIE::DataServer::TDataServer *tdsModelData)
	Construction from a data server containing model parameters.

TGPBuilder *	buildWithDuplicateKandCandCorrFunc ()

void	initVariables ()
	Initialisation of the GPBuilder variables.

void	initMatrices ()
	Initialisation of the matrices.

void	createTrend ()
	Create the matrices for the deterministic trend.

virtual	~TGPBuilder ()
	Default destructor.

Get methods
Bool_t	hasCorrFunction ()
	Return kTRUE if a correlation function is defined, kFALSE otherwise.

Bool_t	hasVarianceInFactor ()
	Return kTRUE if the variance of the GP is determined analytically.

Bool_t	hasMeasurementError ()
	Return kTRUE if a measurement error should be considered to build the GP.

Bool_t	hasTrend ()
	Return kTRUE if a deterministic trend has been defined.

Bool_t	usePrior ()
	Return kTRUE if prior knowledge on the deterministic trend is used (bayesian approach).

int	getReturnOptim ()
	Return the return of the optimisation procedure.

TCorrelationFunction *	getCorrFunction ()
	Return a pointer to the correlation function of the gaussian process.

Int_t	getNInputs ()
	Return the number of input variables.

Int_t	getNObs ()
	Return the number of observations used to build the GP model.

Int_t	getNTrendParams ()
	Return the number of parameters of the deterministic trend.

Double_t *	getObsInputs ()
	Return the observations inputs as a flat array.

Double_t *	getObsOutput ()
	Return the observations outputs.

Double_t *	getObsNormParams ()
	Return the normalisation parameters of the inputs as a flat array.

Double_t *	getCovarianceMatrix ()
	Return the array of the covariance matrix of the gaussian process.

void	setKandCandCorrFunc (Double_t inK, Double_t inC, TCorrelationFunction *corrFunc)

Double_t *	getCorrelationMatrix ()
	Return the array of the correlation matrix of the gaussian process.

Double_t *	getMeasurementErrorCovMatrix ()
	Return the array of the covariance matrix of the measurement errors.

Double_t *	getTrendMatrix ()
	Return the array of the deterministic trend matrix.

TObjArray *	getTrendFormula ()
	Return the deterministic trend list of formulas.

TString	getTrendString ()
	Return the deterministic trend character string.

Double_t *	getTrendPriorMeans ()
	Return the array of the a priori means of the trend parameters.

Double_t *	getTrendPriorCovMat ()
	Return the array of the a priori covariance matrix of the trend parameters.

Double_t	getVariance ()
	Return the variance of the gaussian process.

Double_t	getTolerance ()
	Return the tolerance used to stop the optimisation process

Double_t	getAlpha ()
	Return the alpha parameter, quotient of the variance of the measurement error and the variance of the gaussian process.

Double_t	getRegularisation ()
	Return the regularisation coefficient.

Int_t	getSeed ()
	Return the seed number.

Set methods
void	setLengthRange (double themin, double themax)
	Set the Minimun and maximum boundaries for the correlation length.

void	setLengthMaxScreeningSF (double sf)
	Set the scale factor to reduce the maximum value of the correlation length during screening.

void	setCorrFunction (TCorrelationFunction *corrFunc)
	Set the correlation function of the gaussian process.

void	setCorrFunction (const char *corrFuncType)
	Set the correlation function of the gaussian process from a character string.

void	setVariance (Double_t var)
	Set the variance of the gaussian process.

void	setTolerance (Double_t tol)
	Set the tolerance used to stop the optimisation process.

void	setRegularisation (Double_t regularisation)
	Set the regularisation coefficient.

void	setHasMeasurementError (Bool_t has_measurement_error=kTRUE)
	Indicate if a measurement error should be taken into account or not.

void	setAlpha (Double_t a)
	Set the value of the alpha parameter.

void	setMeasurementErrorCovMatrix (Double_t *covMes)
	Set the covariance matrix of the measurement errors.

void	setMeasurementErrorVariance (Double_t varMes)
	Set the variance of the measurement errors.

void	setTrendString (const char *trend)
	Set the deterministic trend character string.

void	setPriorData (Double_t betaPrior, Double_t qPrior)
	Set the a priori mean and variance of the trend coefficients.

void	setUsePrior (Bool_t use_prior=kTRUE)
	Set wether or not to use a priori knowledge on the deterministic trend parameters.

void	computeCovarianceMatrix (bool print=false)
	Computes the covariance matrix of the observed data.

void	setSeed (Int_t nSeed=0)
	set a seed number

Construction of the Gaussian Process
void	updateObservations ()
	Update the GP builder matrices.

void	exportGPData (const char *fileName)
	Export the minimum information necessary to build a new TKriging object.

void	findOptimalParameters (TString criterion="ML", Int_t screeningSize=100, TString optimAlgo="BFGS", Int_t nbMaxOptimSteps=1000, Bool_t reset=kTRUE, Option_t *option="")
	Search for the optimal parameters of the Gaussian Process.

URANIE::Modeler::TKriging *	buildGP (Bool_t computeLooErrors=kTRUE)
	Build the gaussian process model.

void	cleanAttributeList (TList *inputlist)
	Clean list of attribute.

Printing Log
void	setLog ()

void	unsetLog ()

void	changeLog ()

Bool_t	getLog ()

virtual void	printLog (Option_t *option="")

Public Attributes
URANIE::DataServer::TDataServer *	_tds
	Pointer to the data server containing the observations.

TString	_sInputAttributes
	input attributes names separated by ":"

TString	_sOutputAttributes
	output attribute name

Int_t	_nX
	Number of input attributes.

Int_t	_nY
	Number of output attributes (should be equal to 1)

TCorrelationFunction *	correlationFunction
	The correlation function.

Bool_t	bHas_corrFunction
	kFALSE if correlationFunction == NULL, kTRUE otherwise

Bool_t	bVariance_in_factor
	if kTRUE, GP variance is calculated analytically. Otherwise, it is a parameter of the cost function.

Bool_t	bHas_measurement_error
	if kTRUE, the construction of the GP will take into account the existence of a measurement error variance.

Int_t	_nS
	Number of observations.

Int_t	_nP
	Number of coefficients of the deterministic trend.

Double_t	reg
	regularisation parameter, used to improve numerical stability of matrix operations

Double_t	tolerance
	Tolerance value used to stop the optimisation process.

Double_t	sigma2
	Variance of the gaussian process.

Double_t	alpha
	Quotient of the variance of the measurment error and the variance of the gaussian process.

Double_t *	Vmes
	The covariance matrix of the measurement errors (dimension: _nS * _nS)

Double_t *	H
	Deterministic trend matrix (dimension: _nS * _nP)

Double_t *	C
	The correlation matrix (dimension: _nS * _nS)

Double_t *	K
	The covariance matrix (dimension: _nS * _nS)

Double_t *	xObs
	The matrix of normalised inputs (dimension: _nX * _nS)

Double_t *	yObs
	Array representation of the outputs (size: _nS)

Double_t *	xNormParams
	The matrix of normalisation parameters. Contains min and max value of each input variable: [x1Min, x1Max, x2Min, x2Max,...] (size: _nX * 2)

Double_t	_modlengthMax
	Maximum value of the correlation length if default behaviour is requested to be changed by the user.

Double_t	_modlengthMin
	Minimum value of the correlation length if default behaviour is requested to be changed by the user.

Double_t	_lengthMaxScreeningSF
	Scale factor to change the Maximum value of the correlation length during screening procedure: new length is SF * lengthMax (0 < SF <=1)

TString	trendString
	Character string containing the trend parameters separated by ":" or a trend type ("Const" or "Linear")

TObjArray *	trendCoefList
	List of the formulas of the trend coefficients (_nP elements)

Bool_t	bHas_trend
	kFALSE if no deterministic trend is defined, kTRUE otherwise.

Bool_t	bUse_normalisation
	if kFALSE, the matrix H will be computed with unormalised input data. This happens when the user is providing his own trend function. When the trend is constructed automatically from a trend type, this attribute is set to kTRUE.

Double_t *	BetaPrior
	a priori mean vector for the deterministic trend parameters (size: _nP)

Double_t *	QPrior
	a priori variance matrix for the deterministic trend parameters (size: _nP x _nP)

Bool_t	bUse_prior
	kFALSE if no prior is used, kTRUE otherwise

Int_t	_nSeed
	number of seed for pseudo-random generator

Int_t	_retOptim
	optimisation results

Bool_t	bIsbuildWithDuplicateKandCandCorrFunc
	kTRUE if build with duplicate K, C and CorrFunc, otherwise kFALSE

Private Member Functions
void	LSort (Double_t costValues, Int_t minIndexes, Int_t nbthreads, Int_t n)

Private Attributes
Bool_t	_blog
	Boolean for edit the log.

Detailed Description

Description of the class TGPBuilder.

This class allows to build a gaussian processus (kriging) model from a set of observations and a correlation function. It provides tools to search for the optimal parameters of the correlation function and compute the parameters of the model. It wraps the gpModx functions of the "gaussian process Library" (gpLib), by Jean-Marc Martinez.

Remark about normalisation: the observed input variable are automatically normalised such that they belong to [0,1]. The normalisation coefficients are exported into the TKriging model and new input coordinates are thus automatically normalised. As a consequence, the correlation lengths of the correlation function are normalised as well. A special case happens when the deterministic trend is not a constant or a linear combination of the inputs. In that case, the values of the trend coefficients are calculated in the "original" space.

Constructor & Destructor Documentation

◆ TGPBuilder() [1/4]

URANIE::Modeler::TGPBuilder::TGPBuilder ( )

Default Constructor.

Referenced by ClassImp().

◆ TGPBuilder() [2/4]

URANIE::Modeler::TGPBuilder::TGPBuilder	(	URANIE::DataServer::TDataServer *	tdsObs,
		const char *	varexpinput,
		const char *	varexpoutput,
		TCorrelationFunction *	corrFunc,
		const char *	trend = `""`,
		Option_t *	option = `""`
	)

Construction from a data server, a list of input and output variables and a correlation function object.

Parameters

tdsObs	a pointer to the data server containing the observations.
varexpinput	a character string containing the names of the input variables, separated by colons (":").
varexpoutput	a character string containing the name of the output variable.
corrFunc	a pointer to the correlation function.
trend	a string containing the list of deterministic trend coefficients separated by colons ":" (default = ""). (cf. TGPBuilder::createTrend)
option	a character string containing optional commands. None is avaiblable at this time (default = "").

Exceptions

UErrorException if varexpoutput contains more than one variable name.

◆ TGPBuilder() [3/4]

URANIE::Modeler::TGPBuilder::TGPBuilder	(	URANIE::DataServer::TDataServer *	tdsObs,
		const char *	varexpinput,
		const char *	varexpoutput,
		const char *	corrFunc,
		const char *	trend = `""`,
		Option_t *	option = `""`
	)

Construction from a data server, a list of input and output variables and a correlation function name.

Parameters

tdsObs	a pointer to the data server containing the observations.
varexpinput	a character string containing the names of the input variables, separated by colons (":").
varexpoutput	a character string containing the name of the output variable.
corrFunc	name of the correlation function. Available names are: "gauss", "isogauss", "exponential", "maternI", "maternII", "maternIII", "matern3/2", "matern5/2", "matern7/2". The class name of a correlation function is also valid (e.g. "TExponentialCorrFunction", "TMatern32CorrFunction", etc.)
trend	a string containing the list of deterministic trend coefficients separated by colons ":" (default = ""). (cf. TGPBuilder::createTrend)
option	a character string containing optional commands. None is available at this time (default = "").

Exceptions

UErrorException if varexpoutput contains more than one variable name.

◆ TGPBuilder() [4/4]

URANIE::Modeler::TGPBuilder::TGPBuilder ( URANIE::DataServer::TDataServer * tdsModelData )

Construction from a data server containing model parameters.

This constructor works ONLY with a data server generated by the function TGPBuilder::exportGPData. A set of parameters is written in the #TITLE: field of the file header. These parameters are extracted and used to initialise the GPBuilder. The rest of the file contains the observations (inputs and output). After the construction of the TGPBuilder object, the user can call the function TGPBuilder::buildGP to generate the TKriging object.

Parameters

tdsObs (URANIE::DataServer::TDataServer*): a pointer to the data server containing the observations.

Note: a message is displayed if the user needs to define the measurement error variance or the a priori information on the trend parameters.

◆ ~TGPBuilder()

virtual URANIE::Modeler::TGPBuilder::~TGPBuilder ( )

virtual

Default destructor.

Referenced by ClassImp().

Member Function Documentation

◆ buildGP()

URANIE::Modeler::TKriging * URANIE::Modeler::TGPBuilder::buildGP ( Bool_t computeLooErrors = kTRUE )

Build the gaussian process model.

This function computes the parameters of the gaussian process, necessary to perform new predictions. It then creates a TKriging model and return its pointer.

Note: Calls gpMod0 or gpModh function from gpL

Warning: The user must delete the TKriging object by himself.

Parameters

computeLooErrors (Bool_t): if kTRUE, the Leave One Out errors of the kriging model are computed and stored in the model. (default = kTRUE).

Referenced by ClassImp().

◆ buildWithDuplicateKandCandCorrFunc()

TGPBuilder * URANIE::Modeler::TGPBuilder::buildWithDuplicateKandCandCorrFunc ( )

Referenced by ClassImp(), and ClassImp().

◆ changeLog()

void URANIE::Modeler::TGPBuilder::changeLog ( )

inline

References _blog.

◆ cleanAttributeList()

void URANIE::Modeler::TGPBuilder::cleanAttributeList ( TList * inputlist )

Clean list of attribute.

Referenced by ClassImp().

◆ computeCovarianceMatrix()

void URANIE::Modeler::TGPBuilder::computeCovarianceMatrix ( bool print = false )

Computes the covariance matrix of the observed data.

Referenced by ClassImp(), ClassImp(), ClassImp(), and ClassImp().

◆ createTrend()

void URANIE::Modeler::TGPBuilder::createTrend ( )

Create the matrices for the deterministic trend.

Referenced by ClassImp().

◆ exportGPData()

void URANIE::Modeler::TGPBuilder::exportGPData ( const char * fileName )

Export the minimum information necessary to build a new TKriging object.

This function allows to save into an ASCII file (Salome table format) a set of information and data that can later be read by a TGPBuilder object to recreate the full TKriging object. The saved information is:

the type of the correlation function
the correlation function parameters (not normalised)
the gaussian process variance
if needed, the deterministic trend character string
if needed, the variance of the measurement error (except if it was user defined)
if needed, the regularisation coefficient
the observations inputs and outputs.
Parameters

fileName (const char*): name of the output file.

Referenced by ClassImp().

◆ findOptimalParameters()

void URANIE::Modeler::TGPBuilder::findOptimalParameters	(	TString	criterion = `"ML"`,
		Int_t	screeningSize = `100`,
		TString	optimAlgo = `"BFGS"`,
		Int_t	nbMaxOptimSteps = `1000`,
		Bool_t	reset = `kTRUE`,
		Option_t *	option = `""`
	)

Search for the optimal parameters of the Gaussian Process.

This function allows to search for the parameters of the correlation function which minimise a given criterion. The function works in two steps:

a screening of the parameter space is performed to find a good starting point for the optimisation;
an optimisation process is performed to refine the results.

To skip the screening step, simply set the screening size to 0. The initial point of the optimisation will then be chosen as the set of default value of the parameters to optimise (see below), or their current values if "reset" if set to kFALSE.

To skip the optimisation step, simply set the optimisation algorithm to "" (empty string). The optimal parameters will then be the one corresponding to the optimal output of the screening.

This function is set to be "user-friendly", and takes care of most of the tedious part of the process. This implies some limitations:

hyper-parameters ranges can't be set by the user;
most of the optimisation parameters can't be set by the user;
currently supported correlation functions are: Gaussian, Isotropic Gaussian, Exponential, MaternI, MaternII, MaternIII, Matern3/2, Matern5/2 and Matern7/2.

The consequence of these limitations is that the "optimal" parameters might not be that optimal. Thus, if the resulting model isn't good enough, using the "not so user-friendly" method might be necessary.

Here is the list of the current "hard coded" ranges and default values for the parameters to optimise:

correlation lengths: [1e-3, 10.0], 0.01;
exponent: [0.5, 2.0], 1.0;
Matern's regularity (\nu): [0.5, 7.5], 1.5;
alpha: [1e-6, 10.0], 0.00316;
GP variance coefficient: [1e-6, 1e2], 0.01; (this value is multiplied by the observation output's variance).

Warning: This function throws an exception if the NLOpt optimisation library isn't available.

Parameters

criterion	(TString): criterion to optimize. Can be "ML" (maximum likelyhood), "ReML" (restrained maximum likelyhood), or "LOO" (leave one out cross-validation). "ML" can't be used in bayesian mode, and ReML can't be used without a deterministic trend. (default = "ML").
screeningSize	(Int_t): number of parameters combinaisons to evaluate. An LHS sampling of this size is performed on the parameter space, and evaluated by the cost function. (default = 100).
optimAlgo	(TString): name of the optimisation algorithm. Can be: "NelderMead", "Subplexe", "Cobyla", "Bobyqa", "Praxis", "BFGS", "Newton", "MMA", "SLSQP", "VariableMetric". For details about these algorithms, cf. NLopt website: http://ab-initio.mit.edu/wiki/index.php/NLopt_Algorithms. (default = "BFGS").
nbMaxOptimSteps	(Int_t): maximum number of optimisation steps. This can be used when the cost function is long to evaluate and the optimisation algorithm is slow to converge. (default = 1000).
reset	(Bool_t): if kTRUE, the parameters to optimise are reset to their default values. If kFALSE, the current state of the parameters (correlation lengths, alpha, GP variance, etc.) is used. (default = kTRUE).

Referenced by ClassImp().

◆ getAlpha()

Double_t URANIE::Modeler::TGPBuilder::getAlpha ( )

inline

Return the alpha parameter, quotient of the variance of the measurement error and the variance of the gaussian process.

References alpha.

◆ getCorrelationMatrix()

Double_t * URANIE::Modeler::TGPBuilder::getCorrelationMatrix ( )

inline

Return the array of the correlation matrix of the gaussian process.

References C.

◆ getCorrFunction()

TCorrelationFunction * URANIE::Modeler::TGPBuilder::getCorrFunction ( )

inline

Return a pointer to the correlation function of the gaussian process.

References correlationFunction.

Referenced by ClassImp(), ClassImp(), and ClassImp().

◆ getCovarianceMatrix()

Double_t * URANIE::Modeler::TGPBuilder::getCovarianceMatrix ( )

inline

Return the array of the covariance matrix of the gaussian process.

References K.

◆ getLog()

Bool_t URANIE::Modeler::TGPBuilder::getLog ( )

inline

References _blog.

◆ getMeasurementErrorCovMatrix()

Double_t * URANIE::Modeler::TGPBuilder::getMeasurementErrorCovMatrix ( )

inline

Return the array of the covariance matrix of the measurement errors.

References Vmes.

◆ getNInputs()

Int_t URANIE::Modeler::TGPBuilder::getNInputs ( )

inline

Return the number of input variables.

References _nX.

◆ getNObs()

Int_t URANIE::Modeler::TGPBuilder::getNObs ( )

inline

Return the number of observations used to build the GP model.

References _nS.

◆ getNTrendParams()

Int_t URANIE::Modeler::TGPBuilder::getNTrendParams ( )

inline

Return the number of parameters of the deterministic trend.

References _nP.

◆ getObsInputs()

Double_t * URANIE::Modeler::TGPBuilder::getObsInputs ( )

inline

Return the observations inputs as a flat array.

For M observations of inputs a, b and c, the returned array is organised as follows: [a1, b1, c1, a2, b2, c2, ... , aM, bM, cM]

References xObs.

◆ getObsNormParams()

Double_t * URANIE::Modeler::TGPBuilder::getObsNormParams ( )

inline

Return the normalisation parameters of the inputs as a flat array.

Normalisation parameters are the minimum and maximum values of the input variables. For input variables x1, x2, ... xN, they are stored as follows: [x1min, x1max, x2min, x2max, ... , xNmin, xNmax].

References xNormParams.

◆ getObsOutput()

Double_t * URANIE::Modeler::TGPBuilder::getObsOutput ( )

inline

Return the observations outputs.

References yObs.

◆ getRegularisation()

Double_t URANIE::Modeler::TGPBuilder::getRegularisation ( )

inline

Return the regularisation coefficient.

References reg.

◆ getReturnOptim()

int URANIE::Modeler::TGPBuilder::getReturnOptim ( )

inline

Return the return of the optimisation procedure.

References _retOptim.

◆ getSeed()

Int_t URANIE::Modeler::TGPBuilder::getSeed ( )

inline

Return the seed number.

References _nSeed.

◆ getTolerance()

Double_t URANIE::Modeler::TGPBuilder::getTolerance ( )

inline

Return the tolerance used to stop the optimisation process

References tolerance.

◆ getTrendFormula()

TObjArray * URANIE::Modeler::TGPBuilder::getTrendFormula ( )

inline

Return the deterministic trend list of formulas.

References trendCoefList.

◆ getTrendMatrix()

Double_t * URANIE::Modeler::TGPBuilder::getTrendMatrix ( )

inline

Return the array of the deterministic trend matrix.

References H.

◆ getTrendPriorCovMat()

Double_t * URANIE::Modeler::TGPBuilder::getTrendPriorCovMat ( )

inline

Return the array of the a priori covariance matrix of the trend parameters.

References QPrior.

◆ getTrendPriorMeans()

Double_t * URANIE::Modeler::TGPBuilder::getTrendPriorMeans ( )

inline

Return the array of the a priori means of the trend parameters.

References BetaPrior.

◆ getTrendString()

TString URANIE::Modeler::TGPBuilder::getTrendString ( )

inline

Return the deterministic trend character string.

References trendString.

◆ getVariance()

Double_t URANIE::Modeler::TGPBuilder::getVariance ( )

inline

Return the variance of the gaussian process.

References sigma2.

◆ hasCorrFunction()

Bool_t URANIE::Modeler::TGPBuilder::hasCorrFunction ( )

inline

Return kTRUE if a correlation function is defined, kFALSE otherwise.

References bHas_corrFunction.

Referenced by ClassImp(), ClassImp(), and ClassImp().

◆ hasMeasurementError()

Bool_t URANIE::Modeler::TGPBuilder::hasMeasurementError ( )

inline

Return kTRUE if a measurement error should be considered to build the GP.

References bHas_measurement_error.

Referenced by ClassImp(), ClassImp(), and ClassImp().

◆ hasTrend()

Bool_t URANIE::Modeler::TGPBuilder::hasTrend ( )

inline

Return kTRUE if a deterministic trend has been defined.

References bHas_trend.

Referenced by ClassImp(), and ClassImp().

◆ hasVarianceInFactor()

Bool_t URANIE::Modeler::TGPBuilder::hasVarianceInFactor ( )

inline

Return kTRUE if the variance of the GP is determined analytically.

References bVariance_in_factor.

Referenced by ClassImp(), ClassImp(), and ClassImp().

◆ initMatrices()

void URANIE::Modeler::TGPBuilder::initMatrices ( )

Initialisation of the matrices.

Referenced by ClassImp().

◆ initVariables()

void URANIE::Modeler::TGPBuilder::initVariables ( )

Initialisation of the GPBuilder variables.

Referenced by ClassImp().

◆ LSort()

void URANIE::Modeler::TGPBuilder::LSort	(	Double_t *	costValues,
		Int_t *	minIndexes,
		Int_t	nbthreads,
		Int_t	n
	)

private

Referenced by ClassImp().

◆ printLog()

virtual void URANIE::Modeler::TGPBuilder::printLog ( Option_t * option = "" )

virtual

Referenced by ClassImp().

◆ setAlpha()

void URANIE::Modeler::TGPBuilder::setAlpha ( Double_t a )

inline

Set the value of the alpha parameter.

This parameter is used when the variance of the measurement error can be expressed as sigma2*alpha, where sigma2 is the variance of the GP.

Note: calling this function automatically sets the flags bHas_measurement_error and bVariance_in_factor to kTRUE.

Parameters

a	the value of the parameter alpha

References alpha, bHas_measurement_error, and bVariance_in_factor.

Referenced by ClassImp(), ClassImp(), ClassImp(), and ClassImp().

◆ setCorrFunction() [1/2]

void URANIE::Modeler::TGPBuilder::setCorrFunction ( const char * corrFuncType )

Set the correlation function of the gaussian process from a character string.

The name of the correlation function can either be its class name (e.g. "TExponentialCorrFunction", "TMatern32CorrFunction", etc.), or one of these (case insensitive) keywords: "gauss", "isogauss", "exponential", "maternI", "maternII", "maternIII", "matern3/2", "matern5/2", "matern7/2".

Parameters

corrFunc name of the correlation function.

◆ setCorrFunction() [2/2]

void URANIE::Modeler::TGPBuilder::setCorrFunction ( TCorrelationFunction * corrFunc )

Set the correlation function of the gaussian process.

The internal correlation function is cloned from the correlation function provided as parameter. This means that the later isn't modified by the GPBuilder.

Parameters

corrFunc a pointer to the correlation function of the gaussian process.

Referenced by ClassImp().

◆ setHasMeasurementError()

void URANIE::Modeler::TGPBuilder::setHasMeasurementError ( Bool_t has_measurement_error = kTRUE )

inline

Indicate if a measurement error should be taken into account or not.

This function sets the bHas_measurement_error flag which indicates wether a measurement error should be considered to build the Gaussian Process. If kTRUE, and the user doesn't provide a covariance matrix (using setMeasurementErrorCovMatrix or setMeasurementErrorVariance), the parameter alpha will be a parameter of the cost function, and the variance of the measurment error will be sigma2*alpha.

Parameters

has_measurement_error flag to tell if measurement error is taken into account. (default = kTRUE)

References bHas_measurement_error.

Referenced by ClassImp().

◆ setKandCandCorrFunc()

void URANIE::Modeler::TGPBuilder::setKandCandCorrFunc	(	Double_t *	inK,
		Double_t *	inC,
		TCorrelationFunction *	corrFunc
	)

inline

References bIsbuildWithDuplicateKandCandCorrFunc, C, correlationFunction, and K.

◆ setLengthMaxScreeningSF()

void URANIE::Modeler::TGPBuilder::setLengthMaxScreeningSF ( double sf )

Set the scale factor to reduce the maximum value of the correlation length during screening.

This function changes the value of the scale factor that can be used to reduce the maximum value of the correlation length during the screening procedure .

Parameters

sf	a double (default is 1) that is the scale factor to define a new lengthMax specifically for the screening, as lengthMaxScr = sf * lengthMax. Exception if sf < 0 or sf > 1

Referenced by ClassImp().

◆ setLengthRange()

void URANIE::Modeler::TGPBuilder::setLengthRange	(	double	themin,
		double	themax
	)

Set the Minimun and maximum boundaries for the correlation length.

This function changes the default value of the Minimun and maximum boundaries used to define the space definition of the correlation length.

Parameters

themin	a double that is the min value allowed for the correlation length (default being 1e-3).
themax	a double that is the max value allowed for the correlation length (default being 10).

Referenced by ClassImp().

◆ setLog()

void URANIE::Modeler::TGPBuilder::setLog ( )

inline

References _blog.

◆ setMeasurementErrorCovMatrix()

void URANIE::Modeler::TGPBuilder::setMeasurementErrorCovMatrix ( Double_t * covMes )

Set the covariance matrix of the measurement errors.

The data from covMes are copied into Vmes. covMes MUST be a valid covariance matrix.

Note: Calling this function automatically sets the flag bHas_measurement_error to kTRUE and the flag bVariance_in_factor to kFALSE, meaning that the GP variance is an input parameter of the cost function.

Parameters

covMes "flat" array of nS*nS doubles.

Referenced by ClassImp().

◆ setMeasurementErrorVariance()

void URANIE::Modeler::TGPBuilder::setMeasurementErrorVariance ( Double_t varMes )

Set the variance of the measurement errors.

The diagonal of the matrix Vmes is filled with varMes.

Note: Calling this function automatically sets the flag bHas_measurement_error to kTRUE and the flag bVariance_in_factor to kFALSE, meaning that the GP variance is an input parameter of the cost function.

Parameters

varMes variance of the measurement error.

Referenced by ClassImp().

◆ setPriorData()

void URANIE::Modeler::TGPBuilder::setPriorData	(	Double_t *	betaPrior,
		Double_t *	qPrior
	)

Set the a priori mean and variance of the trend coefficients.

This function should not be used if no trend was defined. It allows to set the a priori mean and variance of the trend coefficients.

Warning: the size of the array is no checked !

Note: Calling this function automatically sets the flag bUse_prior to kTRUE.

Parameters

betaPrior	a priori mean vector for the deterministic trend parameters (size: _nP)
qPrior	a priori covariance matrix for the deterministic trend parameters (size: _nP x _nP)

Exceptions

UranieError if no trend was defined.

Referenced by ClassImp().

◆ setRegularisation()

void URANIE::Modeler::TGPBuilder::setRegularisation ( Double_t regularisation )

inline

Set the regularisation coefficient.

The regularisation coefficient is a small (~ 1e-12 to 1e-6) value that is added to the diagonal of the correlation matrix when it is poorly conditionned. It should only be set when matrix decomposition problems arise.

Parameters

regularisation value of the regularisation coefficient. It is typically between 1e-12 and 1e-6.

References reg.

Referenced by ClassImp().

◆ setSeed()

void URANIE::Modeler::TGPBuilder::setSeed ( Int_t nSeed = 0 )

set a seed number

This function set the value of _nSeed which indicates the seed number to use by the random generator

Parameters

nSeed the seed number to use by the random generator

Referenced by ClassImp().

◆ setTolerance()

void URANIE::Modeler::TGPBuilder::setTolerance ( Double_t tol )

inline

Set the tolerance used to stop the optimisation process.

Parameters

var	value of the tolerance.

References tolerance.

◆ setTrendString()

void URANIE::Modeler::TGPBuilder::setTrendString ( const char * trend )

inline

Set the deterministic trend character string.

This function creates the deterministic trend matrix based on a list of trend coefficients. This list is created from a character string. It can be either a keyword describing the trend type ("const" for a constant value or "linear" for a linear combination of the inputs) or a set of ROOT style formulas of the data server attributes, separated by ":" (cf. ROOT documentation of TFormula::Analyze). For example, if we have attributes x1 and x2, the following strings are valid trends:

"const" will create a constant trend of the form: a0
"linear" will create a trend of the form: a0 + a1*X1 + a2*X2, where X1 et X2 are the normalised values of x1 and x2
"1:x1:x1*x1:log(x2)" will create a trend of the form: a0 + a1*x1 + a2*x1^2 + a3*log(x2)
Warning
if a coefficient formula doesn't contain any reference to an attribute of the data server, then it MUST be a constant expression (like "1", "log(2)", "pi", etc.). Otherwise, an error will occure.

Parameters

trend the character string containing the trend type ("const" or "linear"), or a the list of coefficients separated by colons ":".

References trendString.

◆ setUsePrior()

void URANIE::Modeler::TGPBuilder::setUsePrior ( Bool_t use_prior = kTRUE )

Set wether or not to use a priori knowledge on the deterministic trend parameters.

This function should not be used if no trend and no prior data were defined.

Parameters

use_prior if kTRUE, the a priori knowledge is used (default = kTRUE).

Exceptions

UranieError if no trend was defined.

Referenced by ClassImp().

◆ setVariance()

void URANIE::Modeler::TGPBuilder::setVariance ( Double_t var )

inline

Set the variance of the gaussian process.

Parameters

var	value of the variance.

References sigma2.

Referenced by ClassImp(), ClassImp(), ClassImp(), and ClassImp().

◆ unsetLog()

void URANIE::Modeler::TGPBuilder::unsetLog ( )

inline

References _blog.

◆ updateObservations()

void URANIE::Modeler::TGPBuilder::updateObservations ( )

Update the GP builder matrices.

This function shall be called when the data server containing the observations is modified. This is updating the matrices but doesn't change the GP hyper-parameters (correlation lengths, variance, etc.)

Referenced by ClassImp().

◆ usePrior()

Bool_t URANIE::Modeler::TGPBuilder::usePrior ( )

inline

Return kTRUE if prior knowledge on the deterministic trend is used (bayesian approach).

References bUse_prior.

Referenced by ClassImp().

Member Data Documentation

◆ _blog

Bool_t URANIE::Modeler::TGPBuilder::_blog

private

Boolean for edit the log.

Referenced by changeLog(), ClassImp(), getLog(), setLog(), and unsetLog().

◆ _lengthMaxScreeningSF

Double_t URANIE::Modeler::TGPBuilder::_lengthMaxScreeningSF

Scale factor to change the Maximum value of the correlation length during screening procedure: new length is SF * lengthMax (0 < SF <=1)

Referenced by ClassImp().

◆ _modlengthMax

Double_t URANIE::Modeler::TGPBuilder::_modlengthMax

Maximum value of the correlation length if default behaviour is requested to be changed by the user.

Referenced by ClassImp().

◆ _modlengthMin

Double_t URANIE::Modeler::TGPBuilder::_modlengthMin

Minimum value of the correlation length if default behaviour is requested to be changed by the user.

Referenced by ClassImp().

◆ _nP

Int_t URANIE::Modeler::TGPBuilder::_nP

Number of coefficients of the deterministic trend.

Referenced by ClassImp(), and getNTrendParams().

◆ _nS

Int_t URANIE::Modeler::TGPBuilder::_nS

Number of observations.

Referenced by ClassImp(), and getNObs().

◆ _nSeed

Int_t URANIE::Modeler::TGPBuilder::_nSeed

number of seed for pseudo-random generator

Referenced by ClassImp(), and getSeed().

◆ _nX

Int_t URANIE::Modeler::TGPBuilder::_nX

Number of input attributes.

Referenced by ClassImp(), and getNInputs().

◆ _nY

Int_t URANIE::Modeler::TGPBuilder::_nY

Number of output attributes (should be equal to 1)

Referenced by ClassImp().

◆ _retOptim

Int_t URANIE::Modeler::TGPBuilder::_retOptim

optimisation results

Referenced by ClassImp(), and getReturnOptim().

◆ _sInputAttributes

TString URANIE::Modeler::TGPBuilder::_sInputAttributes

input attributes names separated by ":"

Referenced by ClassImp().

◆ _sOutputAttributes

TString URANIE::Modeler::TGPBuilder::_sOutputAttributes

output attribute name

Referenced by ClassImp().

◆ _tds

URANIE::DataServer::TDataServer* URANIE::Modeler::TGPBuilder::_tds

Pointer to the data server containing the observations.

Referenced by ClassImp().

◆ alpha

Double_t URANIE::Modeler::TGPBuilder::alpha

Quotient of the variance of the measurment error and the variance of the gaussian process.

Referenced by ClassImp(), getAlpha(), and setAlpha().

◆ BetaPrior

Double_t* URANIE::Modeler::TGPBuilder::BetaPrior

a priori mean vector for the deterministic trend parameters (size: _nP)

Referenced by ClassImp(), and getTrendPriorMeans().

◆ bHas_corrFunction

Bool_t URANIE::Modeler::TGPBuilder::bHas_corrFunction

kFALSE if correlationFunction == NULL, kTRUE otherwise

Referenced by ClassImp(), and hasCorrFunction().

◆ bHas_measurement_error

Bool_t URANIE::Modeler::TGPBuilder::bHas_measurement_error

if kTRUE, the construction of the GP will take into account the existence of a measurement error variance.

Referenced by ClassImp(), hasMeasurementError(), setAlpha(), and setHasMeasurementError().

◆ bHas_trend

Bool_t URANIE::Modeler::TGPBuilder::bHas_trend

kFALSE if no deterministic trend is defined, kTRUE otherwise.

Referenced by ClassImp(), and hasTrend().

◆ bIsbuildWithDuplicateKandCandCorrFunc

Bool_t URANIE::Modeler::TGPBuilder::bIsbuildWithDuplicateKandCandCorrFunc

kTRUE if build with duplicate K, C and CorrFunc, otherwise kFALSE

Referenced by ClassImp(), and setKandCandCorrFunc().

◆ bUse_normalisation

Bool_t URANIE::Modeler::TGPBuilder::bUse_normalisation

if kFALSE, the matrix H will be computed with unormalised input data. This happens when the user is providing his own trend function. When the trend is constructed automatically from a trend type, this attribute is set to kTRUE.

Referenced by ClassImp().

◆ bUse_prior

Bool_t URANIE::Modeler::TGPBuilder::bUse_prior

kFALSE if no prior is used, kTRUE otherwise

Referenced by ClassImp(), and usePrior().

◆ bVariance_in_factor

Bool_t URANIE::Modeler::TGPBuilder::bVariance_in_factor

if kTRUE, GP variance is calculated analytically. Otherwise, it is a parameter of the cost function.

Referenced by ClassImp(), hasVarianceInFactor(), and setAlpha().

◆ C

Double_t* URANIE::Modeler::TGPBuilder::C

The correlation matrix (dimension: _nS * _nS)

Referenced by ClassImp(), ClassImp(), getCorrelationMatrix(), and setKandCandCorrFunc().

◆ correlationFunction

TCorrelationFunction* URANIE::Modeler::TGPBuilder::correlationFunction

The correlation function.

Referenced by ClassImp(), getCorrFunction(), and setKandCandCorrFunc().

◆ H

Double_t* URANIE::Modeler::TGPBuilder::H

Deterministic trend matrix (dimension: _nS * _nP)

Referenced by ClassImp(), and getTrendMatrix().

◆ K

Double_t* URANIE::Modeler::TGPBuilder::K

The covariance matrix (dimension: _nS * _nS)

Referenced by ClassImp(), ClassImp(), getCovarianceMatrix(), and setKandCandCorrFunc().

◆ QPrior

Double_t* URANIE::Modeler::TGPBuilder::QPrior

a priori variance matrix for the deterministic trend parameters (size: _nP x _nP)

Referenced by ClassImp(), and getTrendPriorCovMat().

◆ reg

Double_t URANIE::Modeler::TGPBuilder::reg

regularisation parameter, used to improve numerical stability of matrix operations

Referenced by ClassImp(), getRegularisation(), and setRegularisation().

◆ sigma2

Double_t URANIE::Modeler::TGPBuilder::sigma2

Variance of the gaussian process.

Referenced by ClassImp(), getVariance(), and setVariance().

◆ tolerance

Double_t URANIE::Modeler::TGPBuilder::tolerance

Tolerance value used to stop the optimisation process.

Referenced by ClassImp(), getTolerance(), and setTolerance().

◆ trendCoefList

TObjArray* URANIE::Modeler::TGPBuilder::trendCoefList

List of the formulas of the trend coefficients (_nP elements)

Referenced by ClassImp(), and getTrendFormula().

◆ trendString

TString URANIE::Modeler::TGPBuilder::trendString

Character string containing the trend parameters separated by ":" or a trend type ("Const" or "Linear")

Referenced by ClassImp(), getTrendString(), and setTrendString().

◆ Vmes

Double_t* URANIE::Modeler::TGPBuilder::Vmes

The covariance matrix of the measurement errors (dimension: _nS * _nS)

Referenced by ClassImp(), and getMeasurementErrorCovMatrix().

◆ xNormParams

Double_t* URANIE::Modeler::TGPBuilder::xNormParams

The matrix of normalisation parameters. Contains min and max value of each input variable: [x1Min, x1Max, x2Min, x2Max,...] (size: _nX * 2)

Referenced by ClassImp(), and getObsNormParams().

◆ xObs

Double_t* URANIE::Modeler::TGPBuilder::xObs

The matrix of normalised inputs (dimension: _nX * _nS)

Referenced by ClassImp(), and getObsInputs().

◆ yObs

Double_t* URANIE::Modeler::TGPBuilder::yObs

Array representation of the outputs (size: _nS)

Referenced by ClassImp(), and getObsOutput().

Documentation / Manuel développeur

Modules disponibles

Public Member Functions

Public Attributes

Private Member Functions

Private Attributes

Detailed Description

Constructor & Destructor Documentation

◆ TGPBuilder() [1/4]

◆ TGPBuilder() [2/4]

◆ TGPBuilder() [3/4]

◆ TGPBuilder() [4/4]

◆ ~TGPBuilder()

Member Function Documentation

◆ buildGP()

◆ buildWithDuplicateKandCandCorrFunc()

◆ changeLog()

◆ cleanAttributeList()

◆ computeCovarianceMatrix()

◆ createTrend()

◆ exportGPData()

◆ findOptimalParameters()

◆ getAlpha()

◆ getCorrelationMatrix()

◆ getCorrFunction()

◆ getCovarianceMatrix()

◆ getLog()

◆ getMeasurementErrorCovMatrix()

◆ getNInputs()

◆ getNObs()

◆ getNTrendParams()

◆ getObsInputs()

◆ getObsNormParams()

◆ getObsOutput()

◆ getRegularisation()

◆ getReturnOptim()

◆ getSeed()

◆ getTolerance()

◆ getTrendFormula()

◆ getTrendMatrix()

◆ getTrendPriorCovMat()

◆ getTrendPriorMeans()

◆ getTrendString()

◆ getVariance()

◆ hasCorrFunction()

◆ hasMeasurementError()

◆ hasTrend()

◆ hasVarianceInFactor()

◆ initMatrices()

◆ initVariables()

◆ LSort()

◆ printLog()

◆ setAlpha()

◆ setCorrFunction() [1/2]

◆ setCorrFunction() [2/2]

◆ setHasMeasurementError()

◆ setKandCandCorrFunc()

◆ setLengthMaxScreeningSF()

◆ setLengthRange()

◆ setLog()

◆ setMeasurementErrorCovMatrix()

◆ setMeasurementErrorVariance()

◆ setPriorData()

◆ setRegularisation()

◆ setSeed()

◆ setTolerance()

◆ setTrendString()

◆ setUsePrior()

◆ setVariance()

◆ unsetLog()

◆ updateObservations()

◆ usePrior()

Member Data Documentation

◆ _blog

◆ _lengthMaxScreeningSF

◆ _modlengthMax

◆ _modlengthMin

◆ _nP

◆ _nS

◆ _nSeed