madness::QuasiNewton Class Reference

Optimization via quasi-Newton (BFGS or SR1 update) More...

#include <solvers.h>

Inheritance diagram for madness::QuasiNewton:

Collaboration diagram for madness::QuasiNewton:

Public Member Functions
	QuasiNewton (const std::shared_ptr< OptimizationTargetInterface > &tar, int maxiter=20, double tol=1e-6, double value_precision=1e-12, double gradient_precision=1e-12)
virtual	~QuasiNewton ()
bool	converged () const
	After running the optimizer returns true if converged.
double	gradient_norm () const
	Value of gradient norm.
Tensor< double >	new_search_direction (const Tensor< double > &g) const
bool	optimize (Tensor< double > &x)
	Runs the optimizer.
void	reset_hessian ()
	Resets Hessian to default guess.
void	set_hessian (const Tensor< double > &matrix)
	Sets Hessian to given matrix.
void	set_test (const bool &test_level)
	Choose update method (currently only "BFGS" or "SR1")
void	set_update (const std::string &method)
	Choose update method (currently only "BFGS" or "SR1")
double	value () const
	Value of objective function.
Public Member Functions inherited from madness::OptimizerInterface
virtual	~OptimizerInterface ()

Static Public Member Functions
static void	hessian_update_bfgs (const Tensor< double > &dx, const Tensor< double > &dg, Tensor< double > &hessian)
	make this static for other QN classed to have access to it
static void	hessian_update_sr1 (const Tensor< double > &s, const Tensor< double > &y, Tensor< double > &hessian)
	make this static for other QN classed to have access to it
static double	line_search (double a1, double f0, double dxgrad, const Tensor< double > &x, const Tensor< double > &dx, std::shared_ptr< OptimizationTargetInterface > target, double value_precision)
	make this static for other QN classed to have access to it

Protected Attributes
double	f
double	gnorm
const double	gradient_precision
Tensor< double >	h
const int	maxiter
int	n
bool	printtest
std::shared_ptr< OptimizationTargetInterface >	target
const double	tol
std::string	update
const double	value_precision

Detailed Description

Optimization via quasi-Newton (BFGS or SR1 update)

This is presently not a low memory algorithm ... we really need one!

Constructor & Destructor Documentation

QuasiNewton()

QuasiNewton::QuasiNewton	(	const std::shared_ptr< OptimizationTargetInterface > &	tar,
		int	maxiter = 20,
		double	tol = 1e-6,
		double	value_precision = 1e-12,
		double	gradient_precision = 1e-12
	)

References target.

~QuasiNewton()

virtual madness::QuasiNewton::~QuasiNewton ( )

inlinevirtual

Member Function Documentation

converged()

bool QuasiNewton::converged ( ) const

virtual

After running the optimizer returns true if converged.

Returns

True if converged

Implements madness::OptimizerInterface.

References gnorm, and tol.

Referenced by optimize().

gradient_norm()

double QuasiNewton::gradient_norm ( ) const

virtual

Value of gradient norm.

Returns

Norm of gradient of objective function

Implements madness::OptimizerInterface.

References gnorm.

hessian_update_bfgs()

void QuasiNewton::hessian_update_bfgs ( const Tensor< double > &  dx, const Tensor< double > &  dg, Tensor< double > &  hessian  )

static

make this static for other QN classed to have access to it

References std::abs(), madness::BaseTensor::dim(), madness::inner(), n, and madness::nonlinear_vector_solver().

Referenced by optimize(), and madness::MolecularOptimizer::optimize_quasi_newton().

hessian_update_sr1()

void QuasiNewton::hessian_update_sr1 ( const Tensor< double > &  s, const Tensor< double > &  y, Tensor< double > &  hessian  )

static

make this static for other QN classed to have access to it

References std::abs(), e(), madness::inner(), madness::nonlinear_vector_solver(), madness::outer(), q(), and madness::Tensor< T >::scale().

Referenced by optimize(), and madness::MolecularOptimizer::optimize_quasi_newton().

line_search()

double QuasiNewton::line_search ( double  a1, double  f0, double  dxgrad, const Tensor< double > &  x, const Tensor< double > &  dx, std::shared_ptr< OptimizationTargetInterface >  target, double  value_precision  )

static

make this static for other QN classed to have access to it

References a1, a2, std::abs(), f1, madness::nonlinear_vector_solver(), madness::print(), and value_precision.

Referenced by optimize(), and madness::MolecularOptimizer::optimize_quasi_newton().

new_search_direction()

Tensor< double > QuasiNewton::new_search_direction

(

const Tensor< double > &

g

)

const

References std::abs(), e(), madness::g, gradient_precision, h, madness::inner(), n, madness::nonlinear_vector_solver(), printtest, madness::syev(), tol, and v.

Referenced by optimize().

optimize()

bool QuasiNewton::optimize ( Tensor< double > &  x )

virtual

Runs the optimizer.

Returns

True if converged

Implements madness::OptimizerInterface.

References converged(), madness::BaseTensor::dim(), f, madness::g, gnorm, h, hessian_update_bfgs(), hessian_update_sr1(), line_search(), maxiter, n, new_search_direction(), madness::nonlinear_vector_solver(), madness::Tensor< T >::normf(), madness::print(), printtest, madness::Tensor< T >::scale(), madness::BaseTensor::size(), target, update, and value_precision.

reset_hessian()

void madness::QuasiNewton::reset_hessian ( )

inline

Resets Hessian to default guess.

References h.

set_hessian()

void madness::QuasiNewton::set_hessian ( const Tensor< double > &  matrix )

inline

Sets Hessian to given matrix.

References madness::copy(), and h.

set_test()

void QuasiNewton::set_test

(

const bool &

test_level

)

Choose update method (currently only "BFGS" or "SR1")

References madness::nonlinear_vector_solver(), and printtest.

set_update()

void QuasiNewton::set_update

(

const std::string &

method

)

Choose update method (currently only "BFGS" or "SR1")

References update.

value()

double QuasiNewton::value ( ) const

virtual

Value of objective function.

Returns

Value of objective function

Implements madness::OptimizerInterface.

References f.

Member Data Documentation

f

double madness::QuasiNewton::f

protected

Referenced by optimize(), and value().

gnorm

double madness::QuasiNewton::gnorm

protected

Referenced by converged(), gradient_norm(), and optimize().

gradient_precision

const double madness::QuasiNewton::gradient_precision

protected

Referenced by new_search_direction().

h

Tensor<double> madness::QuasiNewton::h

protected

Referenced by new_search_direction(), optimize(), reset_hessian(), and set_hessian().

maxiter

const int madness::QuasiNewton::maxiter

protected

Referenced by optimize().

n

int madness::QuasiNewton::n

protected

Referenced by hessian_update_bfgs(), new_search_direction(), and optimize().

printtest

bool madness::QuasiNewton::printtest

protected

Referenced by new_search_direction(), optimize(), and set_test().

target

std::shared_ptr<OptimizationTargetInterface> madness::QuasiNewton::target

protected

Referenced by QuasiNewton(), and optimize().

tol

const double madness::QuasiNewton::tol

protected

Referenced by converged(), and new_search_direction().

update

std::string madness::QuasiNewton::update

protected

Referenced by optimize(), and set_update().

value_precision

const double madness::QuasiNewton::value_precision

protected

Referenced by line_search(), and optimize().

---

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

• solvers.h
• solvers.cc