QMCWolfeStepLengthSelector Class Reference

Algorithm to determine the step length for a line search optimization so that the step length satisfies the Wolfe conditions. More...

#include <QMCWolfeStepLengthSelector.h>

Inheritance diagram for QMCWolfeStepLengthSelector:

List of all members.

Public Member Functions

double stepLength (QMCObjectiveFunction *function, Array1D< double > &position, Array1D< double > &searchDirection, Array1D< double > &gradient, Array2D< double > &unused, double functionValue)

Calculates the step length to use when performing a line search optimization.

Private Member Functions

double calculateLineSearchObjectiveFunction (Array1D< double > &params, Array1D< double > &searchDirection, double stepLength)

double calculateLineSearchObjectiveFunctionDerivative (Array1D< double > &params, Array1D< double > &searchDirection, double stepLength)

double cubicInterpolateStep (double alphaLo, double alphaHi, double phi_0, double phi_alphaLo, double phi_alphaHi, double phiPrime_0)

Perform a cubic interpolation to choose a step length in the interval [a_lo,a_hi].

double zoom (double alphaLo, double alphaHi, double phi_0, double phi_alphaLo, double phi_alphaHi, double phiPrime_0, Array1D< double > &params, Array1D< double > &searchDirection)

Finds a point in the interval [a_lo, a_hi] that satisfy the strong Wolfe conditions.

double wolfeStepLength (double alphaGuess, Array1D< double > &params, Array1D< double > &searchDirection, Array1D< double > &gradient, double functionValue)

Finds a point between 0 and the maximum step length which satisfies the Wolfe conditions.

Private Attributes

QMCObjectiveFunction * OF

The objective function used in calculating the step length.

Detailed Description

Algorithm to determine the step length for a line search optimization so that the step length satisfies the Wolfe conditions.

This algorithm is described in Nocedal and Wright.

Definition at line 28 of file QMCWolfeStepLengthSelector.h.

Member Function Documentation

double QMCWolfeStepLengthSelector::stepLength	(	QMCObjectiveFunction *	function,
		Array1D< double > &	array1,
		Array1D< double > &	array2,
		Array1D< double > &	array3,
		Array2D< double > &	matrix1,
		double	scalar1
	)			`[virtual]`

Calculates the step length to use when performing a line search optimization.

You will have to look at the individual algorithms to see how the input parameters are defined.

Implements QMCLineSearchStepLengthSelectionAlgorithm.

Definition at line 15 of file QMCWolfeStepLengthSelector.cpp.

References OF, and wolfeStepLength().

double QMCWolfeStepLengthSelector::calculateLineSearchObjectiveFunction	(	Array1D< double > &	params,
		Array1D< double > &	searchDirection,
		double	stepLength
	)			`[private]`

Definition at line 30 of file QMCWolfeStepLengthSelector.cpp.

References QMCObjectiveFunction::evaluate(), QMCObjectiveFunctionResult::getScore(), and OF.

Referenced by calculateLineSearchObjectiveFunctionDerivative(), wolfeStepLength(), and zoom().

double QMCWolfeStepLengthSelector::calculateLineSearchObjectiveFunctionDerivative	(	Array1D< double > &	params,
		Array1D< double > &	searchDirection,
		double	stepLength
	)			`[private]`

Definition at line 42 of file QMCWolfeStepLengthSelector.cpp.

References calculateLineSearchObjectiveFunction().

Referenced by wolfeStepLength(), and zoom().

double QMCWolfeStepLengthSelector::cubicInterpolateStep	(	double	alphaLo,
		double	alphaHi,
		double	phi_0,
		double	phi_alphaLo,
		double	phi_alphaHi,
		double	phiPrime_0
	)			`[private]`

Perform a cubic interpolation to choose a step length in the interval [a_lo,a_hi].

See Nocedal and Wright p 56-58.

Parameters:

	alphaLo	lower bound on the step length.
	alphaHi	upper bound on the step length.
	phi_0	objective function value at the current parameters
	phi_alphaLo	1-D objective function value at the smallest step length.
	phi_alphaHi	1-D objective function value at the largest step length.
	phiPrime_0	derivative of the 1-D objective function at the current parameters.

Returns:: calculated step length

Definition at line 58 of file QMCWolfeStepLengthSelector.cpp.

Referenced by wolfeStepLength(), and zoom().

double QMCWolfeStepLengthSelector::zoom	(	double	alphaLo,
		double	alphaHi,
		double	phi_0,
		double	phi_alphaLo,
		double	phi_alphaHi,
		double	phiPrime_0,
		Array1D< double > &	params,
		Array1D< double > &	searchDirection
	)			`[private]`

Finds a point in the interval [a_lo, a_hi] that satisfy the strong Wolfe conditions.

See Nocedal and Wright p 60.

Parameters:

	alphaLo	lower bound on the step length.
	alphaHi	upper bound on the step length.
	phi_0	objective function value at the current parameters
	phi_alphaLo	1-D objective function value at the smallest step length.
	phi_alphaHi	1-D objective function value at the largest step length.
	phiPrime_0	derivative of the 1-D objective function at the current parameters.
	params	set of parameters to evaluate the function with
	searchDirection	direction the line search is being performed along.

Returns:: step length satisfying the strong Wolfe conditions.

Definition at line 129 of file QMCWolfeStepLengthSelector.cpp.

References calculateLineSearchObjectiveFunction(), calculateLineSearchObjectiveFunctionDerivative(), and cubicInterpolateStep().

Referenced by wolfeStepLength().

double QMCWolfeStepLengthSelector::wolfeStepLength	(	double	alphaGuess,
		Array1D< double > &	params,
		Array1D< double > &	searchDirection,
		Array1D< double > &	gradient,
		double	functionValue
	)			`[private]`

Finds a point between 0 and the maximum step length which satisfies the Wolfe conditions.

See Nocedal and Wright p 59.

Parameters:

	alphaGuess	guess at a step length satisfying the Wolfe conditions.
	params	set of parameters to evaluate the function with
	searchDirection	direction the line search is being performed along.
	gradient	gradient of the objective function.
	functionValue	value of the objective function at the current set of parameters.

Returns:: step length satisfying the strong Wolfe conditions.

Definition at line 188 of file QMCWolfeStepLengthSelector.cpp.

References calculateLineSearchObjectiveFunction(), calculateLineSearchObjectiveFunctionDerivative(), cubicInterpolateStep(), and zoom().

Referenced by stepLength().

Member Data Documentation

QMCObjectiveFunction* QMCWolfeStepLengthSelector::OF [private]

The objective function used in calculating the step length.

Definition at line 43 of file QMCWolfeStepLengthSelector.h.

Referenced by calculateLineSearchObjectiveFunction(), and stepLength().

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


Public Member Functions
double	stepLength (QMCObjectiveFunction *function, Array1D< double > &position, Array1D< double > &searchDirection, Array1D< double > &gradient, Array2D< double > &unused, double functionValue)
	Calculates the step length to use when performing a line search optimization.
Private Member Functions
double	calculateLineSearchObjectiveFunction (Array1D< double > &params, Array1D< double > &searchDirection, double stepLength)
double	calculateLineSearchObjectiveFunctionDerivative (Array1D< double > &params, Array1D< double > &searchDirection, double stepLength)
double	cubicInterpolateStep (double alphaLo, double alphaHi, double phi_0, double phi_alphaLo, double phi_alphaHi, double phiPrime_0)
	Perform a cubic interpolation to choose a step length in the interval [a_lo,a_hi].
double	zoom (double alphaLo, double alphaHi, double phi_0, double phi_alphaLo, double phi_alphaHi, double phiPrime_0, Array1D< double > &params, Array1D< double > &searchDirection)
	Finds a point in the interval [a_lo, a_hi] that satisfy the strong Wolfe conditions.
double	wolfeStepLength (double alphaGuess, Array1D< double > &params, Array1D< double > &searchDirection, Array1D< double > &gradient, double functionValue)
	Finds a point between 0 and the maximum step length which satisfies the Wolfe conditions.
Private Attributes
QMCObjectiveFunction *	OF
	The objective function used in calculating the step length.