Detailed Description

Solves a nonlinear program via the projected gradient descent algorithm.

The gradient is determined by differentiation of the costs, or the user can supply a custom gradient function. The projection step is itself an optimization problem in general, but the user can supply a custom projection function. The user can also specify a specific solver interface to be used to solve the projection problem.

The solver terminates if

the projection step fails to find a feasible solution,
the 2-norm distance between subsequent iterates is less than the user-specified tolerance (see ConvergenceTolOptionName), or
a maximum number of iterations have been run (see MaxIterationsOptionName).

Warning: This feature is considered to be experimental and may change or be removed at any time, without any deprecation notice ahead of time.

#include <drake/solvers/projected_gradient_descent_solver.h>

Public Member Functions
	ProjectedGradientDescentSolver ()
	~ProjectedGradientDescentSolver () final
void	SetCustomGradientFunction (const std::function< Eigen::VectorXd(const Eigen::VectorXd &)> &custom_gradient_function)
	Specify a custom gradient function.
void	SetCustomProjectionFunction (const std::function< bool(const Eigen::VectorXd &, Eigen::VectorXd *)> &custom_projection_function)
	Specify a custom projection function.
void	SetProjectionSolverInterface (const SolverInterface *projection_solver_interface)
	Specify a solver interface to be used when solving the L2 projection onto the feasible set of the MathematicalProgram it's being used to solve.
Does not allow copy, move, or assignment
	ProjectedGradientDescentSolver (const ProjectedGradientDescentSolver &)=delete
ProjectedGradientDescentSolver &	operator= (const ProjectedGradientDescentSolver &)=delete
	ProjectedGradientDescentSolver (ProjectedGradientDescentSolver &&)=delete
ProjectedGradientDescentSolver &	operator= (ProjectedGradientDescentSolver &&)=delete
Public Member Functions inherited from SolverBase
	~SolverBase () override
MathematicalProgramResult	Solve (const MathematicalProgram &prog, const std::optional< Eigen::VectorXd > &initial_guess=std::nullopt, const std::optional< SolverOptions > &solver_options=std::nullopt) const
	Like SolverInterface::Solve(), but the result is a return value instead of an output argument.
void	Solve (const MathematicalProgram &, const std::optional< Eigen::VectorXd > &, const std::optional< SolverOptions > &, MathematicalProgramResult *) const override
	Solves an optimization program with optional initial guess and solver options.
bool	available () const override
	Returns true iff support for this solver has been compiled into Drake.
bool	enabled () const override
	Returns true iff this solver is properly configured for use at runtime.
SolverId	solver_id () const final
	Returns the identifier of this solver.
bool	AreProgramAttributesSatisfied (const MathematicalProgram &) const override
	Returns true iff the program's attributes are compatible with this solver's capabilities.
std::string	ExplainUnsatisfiedProgramAttributes (const MathematicalProgram &) const override
	Describes the reasons (if any) why the program is incompatible with this solver's capabilities.
	SolverBase (const SolverBase &)=delete
SolverBase &	operator= (const SolverBase &)=delete
	SolverBase (SolverBase &&)=delete
SolverBase &	operator= (SolverBase &&)=delete
Public Member Functions inherited from SolverInterface
virtual	~SolverInterface ()
	SolverInterface (const SolverInterface &)=delete
SolverInterface &	operator= (const SolverInterface &)=delete
	SolverInterface (SolverInterface &&)=delete
SolverInterface &	operator= (SolverInterface &&)=delete

Static Public Member Functions
static std::string	ConvergenceTolOptionName ()
static std::string	MaxIterationsOptionName ()
static std::string	BacktrackingCOptionName ()
static std::string	BacktrackingTauOptionName ()
static std::string	BacktrackingAlpha0OptionName ()

Static versions of the instance methods with similar names.
static constexpr double	kDefaultConvergenceTol = 1e-12
static constexpr int	kDefaultMaxIterations = 1000
static constexpr double	kDefaultBacktrackingC = 0.5
static constexpr double	kDefaultBacktrackingTau = 0.5
static constexpr double	kDefaultBacktrackingAlpha0 = 0.1
static constexpr int	kDefaultMaxLineSearchSteps = 100
MathematicalProgramResult	Solve (const MathematicalProgram &prog, const std::optional< Eigen::VectorXd > &initial_guess=std::nullopt, const std::optional< SolverOptions > &solver_options=std::nullopt) const
	Like SolverInterface::Solve(), but the result is a return value instead of an output argument.
void	Solve (const MathematicalProgram &, const std::optional< Eigen::VectorXd > &, const std::optional< SolverOptions > &, MathematicalProgramResult *) const override
	Solves an optimization program with optional initial guess and solver options.
static SolverId	id ()
static bool	is_available ()
static bool	is_enabled ()
static bool	ProgramAttributesSatisfied (const MathematicalProgram &)

Additional Inherited Members
Protected Member Functions inherited from SolverBase
	SolverBase (const SolverId &id, std::function< bool()> available, std::function< bool()> enabled, std::function< bool(const MathematicalProgram &)> are_satisfied, std::function< std::string(const MathematicalProgram &)> explain_unsatisfied=nullptr)
	Constructs a SolverBase with the given default implementations of the solver_id(), available(), enabled(), AreProgramAttributesSatisfied(), and ExplainUnsatisfiedProgramAttributes() methods.
virtual void	DoSolve (const MathematicalProgram &prog, const Eigen::VectorXd &initial_guess, const SolverOptions &merged_options, MathematicalProgramResult *result) const
	Hook for subclasses to implement Solve.
Protected Member Functions inherited from SolverInterface
	SolverInterface ()

Constructor & Destructor Documentation

◆ ProjectedGradientDescentSolver() [1/3]

ProjectedGradientDescentSolver ( const ProjectedGradientDescentSolver & )

delete

◆ ProjectedGradientDescentSolver() [2/3]

ProjectedGradientDescentSolver ( ProjectedGradientDescentSolver && )

delete

◆ ProjectedGradientDescentSolver() [3/3]

ProjectedGradientDescentSolver ( )

◆ ~ProjectedGradientDescentSolver()

~ProjectedGradientDescentSolver ( )

final

Member Function Documentation

◆ BacktrackingAlpha0OptionName()

std::string BacktrackingAlpha0OptionName ( )

static

Returns: string key for SolverOptions to set the value of alpha_0 to use for the backtracking line search. Must be positive.

◆ BacktrackingCOptionName()

std::string BacktrackingCOptionName ( )

static

Returns: string key for SolverOptions to set the value of c to use for the backtracking line search. Must be between 0 and 1.

◆ BacktrackingTauOptionName()

std::string BacktrackingTauOptionName ( )

static

Returns: string key for SolverOptions to set the value of tau to use for the backtracking line search. Must be between 0 and 1.

◆ ConvergenceTolOptionName()

std::string ConvergenceTolOptionName ( )

static

Returns: string key for SolverOptions to set the threshold used to determine convergence. It must be positive.

◆ id()

SolverId id ( )

static

◆ is_available()

bool is_available ( )

static

◆ is_enabled()

bool is_enabled ( )

static

◆ MaxIterationsOptionName()

std::string MaxIterationsOptionName ( )

static

Returns: string key for SolverOptions to set the maximum number of iterations. It must be a positive integer.

◆ operator=() [1/2]

ProjectedGradientDescentSolver & operator= ( const ProjectedGradientDescentSolver & )

delete

◆ operator=() [2/2]

ProjectedGradientDescentSolver & operator= ( ProjectedGradientDescentSolver && )

delete

◆ ProgramAttributesSatisfied()

bool ProgramAttributesSatisfied ( const MathematicalProgram & )

static

◆ SetCustomGradientFunction()

void SetCustomGradientFunction ( const std::function< Eigen::VectorXd(const Eigen::VectorXd &)> & custom_gradient_function )

Specify a custom gradient function.

Otherwise, this solver will differentiate through the costs in the MathematicalProgram it's used to solve.

◆ SetCustomProjectionFunction()

void SetCustomProjectionFunction ( const std::function< bool(const Eigen::VectorXd &, Eigen::VectorXd *)> & custom_projection_function )

Specify a custom projection function.

Otherwise, this solver will attempt to solve the L2 projection onto the feasible set of the MathematicalProgram it's used to solve. The projection function should return a boolean value indicating success or failure. It should take in two arguments: the point we are trying to stay close to, and an output argument where the projected value will be placed.

◆ SetProjectionSolverInterface()

void SetProjectionSolverInterface ( const SolverInterface * projection_solver_interface )

Specify a solver interface to be used when solving the L2 projection onto the feasible set of the MathematicalProgram it's being used to solve.

◆ Solve() [1/2]

void Solve	(	const MathematicalProgram &	prog,
		const std::optional< Eigen::VectorXd > &	initial_guess,
		const std::optional< SolverOptions > &	solver_options,
		MathematicalProgramResult *	result ) const

overridevirtual

Solves an optimization program with optional initial guess and solver options.

Note that these initial guess and solver options are not written to prog. If the prog has set an initial guess, and initial_guess is set, then initial_guess takes priority. If the prog has set an option for a solver, and solver_options contains a different value for the same option on the same solver, then solver_options takes priority. Derived implementations of this interface may elect to throw std::exception for badly formed programs.

Implements SolverInterface.

◆ Solve() [2/2]

MathematicalProgramResult Solve	(	const MathematicalProgram &	prog,
		const std::optional< Eigen::VectorXd > &	initial_guess = std::nullopt,
		const std::optional< SolverOptions > &	solver_options = std::nullopt ) const

Like SolverInterface::Solve(), but the result is a return value instead of an output argument.

Member Data Documentation

◆ kDefaultBacktrackingAlpha0

double kDefaultBacktrackingAlpha0 = 0.1

staticconstexpr

◆ kDefaultBacktrackingC

double kDefaultBacktrackingC = 0.5

staticconstexpr

◆ kDefaultBacktrackingTau

double kDefaultBacktrackingTau = 0.5

staticconstexpr

◆ kDefaultConvergenceTol

double kDefaultConvergenceTol = 1e-12

staticconstexpr

◆ kDefaultMaxIterations

int kDefaultMaxIterations = 1000

staticconstexpr

◆ kDefaultMaxLineSearchSteps

int kDefaultMaxLineSearchSteps = 100

staticconstexpr

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

drake/solvers/projected_gradient_descent_solver.h

Detailed Description

Public Member Functions

Static Public Member Functions

Static versions of the instance methods with similar names.

Additional Inherited Members

Constructor & Destructor Documentation

◆ ProjectedGradientDescentSolver() [1/3]

◆ ProjectedGradientDescentSolver() [2/3]

◆ ProjectedGradientDescentSolver() [3/3]

◆ ~ProjectedGradientDescentSolver()

Member Function Documentation

◆ BacktrackingAlpha0OptionName()

◆ BacktrackingCOptionName()

◆ BacktrackingTauOptionName()

◆ ConvergenceTolOptionName()

◆ id()

◆ is_available()

◆ is_enabled()

◆ MaxIterationsOptionName()

◆ operator=() [1/2]

◆ operator=() [2/2]

◆ ProgramAttributesSatisfied()

◆ SetCustomGradientFunction()

◆ SetCustomProjectionFunction()

◆ SetProjectionSolverInterface()

◆ Solve() [1/2]

◆ Solve() [2/2]

Member Data Documentation

◆ kDefaultBacktrackingAlpha0

◆ kDefaultBacktrackingC

◆ kDefaultBacktrackingTau

◆ kDefaultConvergenceTol

◆ kDefaultMaxIterations

◆ kDefaultMaxLineSearchSteps