Provides the interface to compute optimization steps with line search. More...

#include <ROL_LineSearchStep.hpp>

Inheritance diagram for ROL::LineSearchStep< Real >:

Public Member Functions
	LineSearchStep (ROL::ParameterList &parlist, const ROL::Ptr< LineSearch< Real > > &lineSearch=ROL::nullPtr, const ROL::Ptr< Secant< Real > > &secant=ROL::nullPtr, const ROL::Ptr< Krylov< Real > > &krylov=ROL::nullPtr, const ROL::Ptr< NonlinearCG< Real > > &nlcg=ROL::nullPtr)
	Constructor. More...

void	initialize (Vector< Real > &x, const Vector< Real > &s, const Vector< Real > &g, Objective< Real > &obj, BoundConstraint< Real > &bnd, AlgorithmState< Real > &algo_state)
	Initialize step with bound constraint. More...

void	compute (Vector< Real > &s, const Vector< Real > &x, Objective< Real > &obj, BoundConstraint< Real > &bnd, AlgorithmState< Real > &algo_state)
	Compute step. More...

void	update (Vector< Real > &x, const Vector< Real > &s, Objective< Real > &obj, BoundConstraint< Real > &bnd, AlgorithmState< Real > &algo_state)
	Update step, if successful. More...

std::string	printHeader (void) const
	Print iterate header. More...

std::string	printName (void) const
	Print step name. More...

std::string	print (AlgorithmState< Real > &algo_state, bool print_header=false) const
	Print iterate status. More...

Public Member Functions inherited from ROL::Step< Real >
virtual	~Step ()

	Step (void)

virtual void	initialize (Vector< Real > &x, const Vector< Real > &g, Objective< Real > &obj, BoundConstraint< Real > &con, AlgorithmState< Real > &algo_state)
	Initialize step with bound constraint. More...

virtual void	initialize (Vector< Real > &x, const Vector< Real > &g, Vector< Real > &l, const Vector< Real > &c, Objective< Real > &obj, Constraint< Real > &con, AlgorithmState< Real > &algo_state)
	Initialize step with equality constraint. More...

virtual void	initialize (Vector< Real > &x, const Vector< Real > &g, Vector< Real > &l, const Vector< Real > &c, Objective< Real > &obj, Constraint< Real > &con, BoundConstraint< Real > &bnd, AlgorithmState< Real > &algo_state)
	Initialize step with equality constraint. More...

virtual void	compute (Vector< Real > &s, const Vector< Real > &x, const Vector< Real > &l, Objective< Real > &obj, Constraint< Real > &con, AlgorithmState< Real > &algo_state)
	Compute step (equality constraints). More...

virtual void	update (Vector< Real > &x, Vector< Real > &l, const Vector< Real > &s, Objective< Real > &obj, Constraint< Real > &con, AlgorithmState< Real > &algo_state)
	Update step, if successful (equality constraints). More...

virtual void	compute (Vector< Real > &s, const Vector< Real > &x, const Vector< Real > &l, Objective< Real > &obj, Constraint< Real > &con, BoundConstraint< Real > &bnd, AlgorithmState< Real > &algo_state)
	Compute step (equality constraints). More...

virtual void	update (Vector< Real > &x, Vector< Real > &l, const Vector< Real > &s, Objective< Real > &obj, Constraint< Real > &con, BoundConstraint< Real > &bnd, AlgorithmState< Real > &algo_state)
	Update step, if successful (equality constraints). More...

const ROL::Ptr< const StepState< Real > >	getStepState (void) const
	Get state for step object. More...

void	reset (const Real searchSize=1.0)
	Get state for step object. More...

Private Member Functions
Real	GradDotStep (const Vector< Real > &g, const Vector< Real > &s, const Vector< Real > &x, BoundConstraint< Real > &bnd, Real eps=0)

Private Attributes
ROL::Ptr< Step< Real > >	desc_
	Unglobalized step object. More...

ROL::Ptr< Secant< Real > >	secant_
	Secant object (used for quasi-Newton) More...

ROL::Ptr< Krylov< Real > >	krylov_
	Krylov solver object (used for inexact Newton) More...

ROL::Ptr< NonlinearCG< Real > >	nlcg_
	Nonlinear CG object (used for nonlinear CG) More...

ROL::Ptr< LineSearch< Real > >	lineSearch_
	Line-search object. More...

ROL::Ptr< Vector< Real > >	d_

ELineSearch	els_
	enum determines type of line search More...

ECurvatureCondition	econd_
	enum determines type of curvature condition More...

bool	acceptLastAlpha_
	For backwards compatibility. When max function evaluations are reached take last step. More...

bool	usePreviousAlpha_
	If true, use the previously accepted step length (if any) as the new initial step length. More...

int	verbosity_

bool	computeObj_

Real	fval_

ROL::ParameterList	parlist_

std::string	lineSearchName_

Additional Inherited Members
Protected Member Functions inherited from ROL::Step< Real >
ROL::Ptr< StepState< Real > >	getState (void)

Detailed Description

template<class Real>
class ROL::LineSearchStep< Real >

Provides the interface to compute optimization steps with line search.

Suppose \(\mathcal{X}\) is a Hilbert space of functions mapping \(\Xi\) to \(\mathbb{R}\). For example, \(\Xi\subset\mathbb{R}^n\) and \(\mathcal{X}=L^2(\Xi)\) or \(\Xi = \{1,\ldots,n\}\) and \(\mathcal{X}=\mathbb{R}^n\). We assume \(f:\mathcal{X}\to\mathbb{R}\) is twice-continuously Fréchet differentiable and \(a,\,b\in\mathcal{X}\) with \(a\le b\) almost everywhere in \(\Xi\). Note that these line-search algorithms will also work with secant approximations of the Hessian. This step applies to unconstrained and bound constrained optimization problems,

\[ \min_x\quad f(x) \qquad\text{and}\qquad \min_x\quad f(x)\quad\text{s.t.}\quad a\le x\le b, \]

respectively.

For unconstrained problems, given the \(k\)-th iterate \(x_k\) and a descent direction \(s_k\), the line search approximately minimizes the 1D objective function \(\phi_k(t) = f(x_k + t s_k)\). The approximate minimizer \(t\) must satisfy sufficient decrease and curvature conditions into guarantee global convergence. The sufficient decrease condition (often called the Armijo condition) is

\[ \phi_k(t) \le \phi_k(0) + c_1 t \phi_k'(0) \qquad\iff\qquad f(x_k+ts_k) \le f(x_k) + c_1 t \langle \nabla f(x_k),s_k\rangle_{\mathcal{X}} \]

where \(0 < c_1 < 1\). The curvature conditions implemented in ROL include:

Name	Condition	Parameters
Wolfe	\(\phi_k'(t) \ge c_2\phi_k'(0)\)	\(c_1<c_2<1\)
Strong Wolfe	\(\left\|\phi_k'(t)\right\| \le c_2 \left\|\phi_k'(0)\right\|\)	\(c_1<c_2<1\)
Generalized Wolfe	\(c_2\phi_k'(0)\le \phi_k'(t)\le-c_3\phi_k'(0)\)	\(0<c_3<1\)
Approximate Wolfe	\(c_2\phi_k'(0)\le \phi_k'(t)\le(2c_1-1)\phi_k'(0)\)	\(c_1<c_2<1\)
Goldstein	\(\phi_k(0)+(1-c_1)t\phi_k'(0)\le \phi_k(t)\)	\(0<c_1<\frac{1}{2}\)

Note that \(\phi_k'(t) = \langle \nabla f(x_k+ts_k),s_k\rangle_{\mathcal{X}}\).

For bound constrained problems, the line-search step performs a projected search. That is, the line search approximately minimizes the 1D objective function \(\phi_k(t) = f(P_{[a,b]}(x_k+ts_k))\) where \(P_{[a,b]}\) denotes the projection onto the upper and lower bounds. Such line-search algorithms correspond to projected gradient and projected Newton-type algorithms.

For projected methods, we require the notion of an active set of an iterate \(x_k\),

\[ \mathcal{A}_k = \{\, \xi\in\Xi\,:\,x_k(\xi) = a(\xi)\,\}\cap \{\, \xi\in\Xi\,:\,x_k(\xi) = b(\xi)\,\}. \]

Given \(\mathcal{A}_k\) and a search direction \(s_k\), we define the binding set as

\[ \mathcal{B}_k = \{\, \xi\in\Xi\,:\,x_k(\xi) = a(\xi) \;\text{and}\; s_k(\xi) < 0 \,\}\cap \{\, \xi\in\Xi\,:\,x_k(\xi) = b(\xi) \;\text{and}\; s_k(\xi) > 0 \,\}. \]

The binding set contains the values of \(\xi\in\Xi\) such that if \(x_k(\xi)\) is on a bound, then \((x_k+s_k)(\xi)\) will violate bound. Using these definitions, we can redefine the sufficient decrease and curvature conditions from the unconstrained case to the case of bound constraints.

LineSearchStep implements a number of algorithms for both bound constrained and unconstrained optimization. These algorithms are: Steepest descent; Nonlinear CG; Quasi-Newton methods; Inexact Newton methods; Newton's method. These methods are chosen through the EDescent enum.

Definition at line 104 of file ROL_LineSearchStep.hpp.

Constructor & Destructor Documentation

template<class Real>

ROL::LineSearchStep< Real >::LineSearchStep	(	ROL::ParameterList &	parlist,
		const ROL::Ptr< LineSearch< Real > > &	lineSearch = `ROL::nullPtr`,
		const ROL::Ptr< Secant< Real > > &	secant = `ROL::nullPtr`,
		const ROL::Ptr< Krylov< Real > > &	krylov = `ROL::nullPtr`,
		const ROL::Ptr< NonlinearCG< Real > > &	nlcg = `ROL::nullPtr`
	)

inline

Constructor.

Standard constructor to build a LineSearchStep object. Algorithmic specifications are passed in through a ROL::ParameterList. The line-search type, secant type, Krylov type, or nonlinear CG type can be set using user-defined objects.

Parameters

[in]	parlist	is a parameter list containing algorithmic specifications
[in]	lineSearch	is a user-defined line search object
[in]	secant	is a user-defined secant object
[in]	krylov	is a user-defined Krylov object
[in]	nlcg	is a user-defined Nonlinear CG object

Definition at line 171 of file ROL_LineSearchStep.hpp.

References ROL::LineSearchStep< Real >::acceptLastAlpha_, ROL::LineSearchStep< Real >::computeObj_, ROL::LineSearchStep< Real >::econd_, ROL::LineSearchStep< Real >::els_, ROL::LineSearchStep< Real >::lineSearch_, ROL::LineSearchStep< Real >::lineSearchName_, ROL::StringToECurvatureCondition(), ROL::StringToELineSearch(), and ROL::LineSearchStep< Real >::verbosity_.

Member Function Documentation

template<class Real>

Real ROL::LineSearchStep< Real >::GradDotStep	(	const Vector< Real > &	g,
		const Vector< Real > &	s,
		const Vector< Real > &	x,
		BoundConstraint< Real > &	bnd,
		Real	eps = `0`
	)

inlineprivate

Definition at line 130 of file ROL_LineSearchStep.hpp.

References ROL::LineSearchStep< Real >::d_, ROL::Vector< Real >::dot(), ROL::Vector< Real >::dual(), ROL::BoundConstraint< Real >::isActivated(), ROL::BoundConstraint< Real >::project(), ROL::BoundConstraint< Real >::pruneActive(), and ROL::BoundConstraint< Real >::pruneInactive().

Referenced by ROL::LineSearchStep< Real >::compute().

template<class Real>

void ROL::LineSearchStep< Real >::initialize	(	Vector< Real > &	x,
		const Vector< Real > &	s,
		const Vector< Real > &	g,
		Objective< Real > &	obj,
		BoundConstraint< Real > &	con,
		AlgorithmState< Real > &	algo_state
	)

inlinevirtual

Initialize step with bound constraint.

Reimplemented from ROL::Step< Real >.

Definition at line 200 of file ROL_LineSearchStep.hpp.

Referenced by main().

template<class Real>

void ROL::LineSearchStep< Real >::compute	(	Vector< Real > &	s,
		const Vector< Real > &	x,
		Objective< Real > &	obj,
		BoundConstraint< Real > &	bnd,
		AlgorithmState< Real > &	algo_state
	)

inlinevirtual

Compute step.

Computes a trial step, \(s_k\) as defined by the enum EDescent. Once the trial step is determined, this function determines an approximate minimizer of the 1D function \(\phi_k(t) = f(x_k+ts_k)\). This approximate minimizer must satisfy sufficient decrease and curvature conditions.

Parameters

[out]	s	is the computed trial step
[in]	x	is the current iterate
[in]	obj	is the objective function
[in]	bnd	are the bound constraints
[in]	algo_state	contains the current state of the algorithm

Reimplemented from ROL::Step< Real >.

Definition at line 284 of file ROL_LineSearchStep.hpp.

References ROL::LineSearchStep< Real >::acceptLastAlpha_, ROL::Vector< Real >::axpy(), ROL::LineSearchStep< Real >::desc_, ROL::LineSearchStep< Real >::fval_, ROL::Step< Real >::getState(), ROL::AlgorithmState< Real >::gnorm, ROL::LineSearchStep< Real >::GradDotStep(), ROL::BoundConstraint< Real >::isActivated(), ROL::LineSearchStep< Real >::lineSearch_, ROL::Vector< Real >::plus(), ROL::BoundConstraint< Real >::project(), ROL::Vector< Real >::scale(), ROL::Vector< Real >::set(), ROL::AlgorithmState< Real >::value, and zero.

Referenced by main().

template<class Real>

void ROL::LineSearchStep< Real >::update	(	Vector< Real > &	x,
		const Vector< Real > &	s,
		Objective< Real > &	obj,
		BoundConstraint< Real > &	bnd,
		AlgorithmState< Real > &	algo_state
	)

inlinevirtual

Update step, if successful.

Given a trial step, \(s_k\), this function updates \(x_{k+1}=x_k+s_k\). This function also updates the secant approximation.

Parameters

[in,out]	x	is the updated iterate
[in]	s	is the computed trial step
[in]	obj	is the objective function
[in]	con	are the bound constraints
[in]	algo_state	contains the current state of the algorithm