ModuleBase::Opt_TN Class Reference

A class designed to deal with optimization problems with Truncated-Newton (TN) method. At each step, the optimization direction d is determined by roughly solving linear equation Hd=-g, where H is Hessian matrix of f(x), and g is the gradient df(x)/dx. In this section, we get Hd by interpolation, and solve Hd=-g with CG method. We set three truncated conditions: 1) the residual of CG method has decreased more than 90%; 2) the number of CG iteration is more than 50; 3) the residual of CG method doesn't decrease and the CG iteration is larger than 9. More...

#include <opt_TN.hpp>

Collaboration diagram for ModuleBase::Opt_TN:

Public Member Functions
	Opt_TN ()
	~Opt_TN ()
void	allocate (int nx)
	Allocate the space for the arrays in cg_.
void	set_para (double dV)
void	refresh (int nx_new=0)
	Refresh the class. If nx changes, reallocate space in cg_.
template<class T >
void	next_direct (double *px, double *pgradient, int &flag, double *rdirect, T *t, void(T::*p_calGradient)(double *ptemp_x, double *rtemp_gradient))
int	get_iter ()
template<class T >
void	next_direct (double *px, double *pgradient, int &flag, double *rdirect, T *t, void(T::*p_calGradient)(double *px, double *rgradient))
	Get the next direction d with TN method.

Private Member Functions
double	inner_product (double *pa, double *pb, int length)
double	get_epsilon (double *px, double *pcg_direction)
	Get epsilon used in interpolation. epsilon = 2*sqrt(mach_prec_) * (1+|x|) / |d|. || means modulu.

Private Attributes
Opt_CG	cg_
double	dV_ = 1.
int	nx_ = 0
int	iter_ = 0
double	mach_prec_ = 0.

Detailed Description

A class designed to deal with optimization problems with Truncated-Newton (TN) method. At each step, the optimization direction d is determined by roughly solving linear equation Hd=-g, where H is Hessian matrix of f(x), and g is the gradient df(x)/dx. In this section, we get Hd by interpolation, and solve Hd=-g with CG method. We set three truncated conditions: 1) the residual of CG method has decreased more than 90%; 2) the number of CG iteration is more than 50; 3) the residual of CG method doesn't decrease and the CG iteration is larger than 9.

Author

sunliang

Constructor & Destructor Documentation

Opt_TN()

ModuleBase::Opt_TN::Opt_TN ( )

inline

~Opt_TN()

ModuleBase::Opt_TN::~Opt_TN ( )

inline

Member Function Documentation

allocate()

void ModuleBase::Opt_TN::allocate ( int  nx )

inline

Allocate the space for the arrays in cg_.

Parameters

nx	length of the solution array x

Here is the call graph for this function:

Here is the caller graph for this function:

get_epsilon()

double ModuleBase::Opt_TN::get_epsilon ( double *  px, double *  pcg_direction  )

inlineprivate

Get epsilon used in interpolation. epsilon = 2*sqrt(mach_prec_) * (1+|x|) / |d|. || means modulu.

Parameters

px	x
pcg_direction	the direction of cg_

Returns

epsilon

Here is the call graph for this function:

Here is the caller graph for this function:

get_iter()

int ModuleBase::Opt_TN::get_iter ( )

inline

inner_product()

double ModuleBase::Opt_TN::inner_product ( double *  pa, double *  pb, int  length  )

inlineprivate

Here is the call graph for this function:

Here is the caller graph for this function:

next_direct() [1/2]

template<class T >

void ModuleBase::Opt_TN::next_direct	(	double *	px,
		double *	pgradient,
		int &	flag,
		double *	rdirect,
		T *	t,
		void(T::*)(double *ptemp_x, double *rtemp_gradient)	p_calGradient
	)

Here is the caller graph for this function:

next_direct() [2/2]

template<class T >

void ModuleBase::Opt_TN::next_direct	(	double *	px,
		double *	pgradient,
		int &	flag,
		double *	rdirect,
		T *	t,
		void(T::*)(double *px, double *rgradient)	p_calGradient
	)

Get the next direction d with TN method.

Template Parameters

T	for example, ESolver_OF

Parameters

[in]	px	current x.
[in]	pgradient	df(x)/dx.
[out]	flag	record which truncated condition was triggered, 0 for cond.1, 1 for cond.2, and 2 for cond.3.
[out]	rdirect	next optimization direction.
[in]	t	pointer of class T, which contains the gradient function.
[in]	p_calGradient	a function pointer, which calculates gradient at provided x.

Here is the call graph for this function:

refresh()

void ModuleBase::Opt_TN::refresh ( int  nx_new = 0 )

inline

Refresh the class. If nx changes, reallocate space in cg_.

Parameters

nx_new

length of new x, default 0 means the length doesn't change

Here is the call graph for this function:

Here is the caller graph for this function:

set_para()

void ModuleBase::Opt_TN::set_para ( double  dV )

inline

Here is the call graph for this function:

Here is the caller graph for this function:

Member Data Documentation

cg_

Opt_CG ModuleBase::Opt_TN::cg_

private

dV_

double ModuleBase::Opt_TN::dV_ = 1.

private

iter_

int ModuleBase::Opt_TN::iter_ = 0

private

mach_prec_

double ModuleBase::Opt_TN::mach_prec_ = 0.

private

nx_

int ModuleBase::Opt_TN::nx_ = 0

private

---

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

• /home/runner/work/abacus-develop/abacus-develop/source/source_base/opt_TN.hpp