|
|
|
@@ -0,0 +1,744 @@
|
|
|
|
|
#ifndef MFEM_FUNCTIONAL_HPP
|
|
|
|
|
#define MFEM_FUNCTIONAL_HPP
|
|
|
|
|
|
|
|
|
|
#include "../config/config.hpp"
|
|
|
|
|
|
|
|
|
|
#ifdef MFEM_USE_MPI
|
|
|
|
|
#include "../general/communication.hpp"
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
#include "operator.hpp"
|
|
|
|
|
#include "blockvector.hpp"
|
|
|
|
|
#include "solvers.hpp"
|
|
|
|
|
#include <cxxabi.h>
|
|
|
|
|
#include <vector>
|
|
|
|
|
|
|
|
|
|
namespace mfem
|
|
|
|
|
{
|
|
|
|
|
/// @brief A base class for functionals F:R^n->R
|
|
|
|
|
///
|
|
|
|
|
/// This class provides an interface for evaluating
|
|
|
|
|
/// $ F:R^n->R, \nabla F:R^n->R^n $, and $ \nabla^2 F:R^n x R^n ->R^n $.
|
|
|
|
|
/// F.Mult(x, y) evaluates the functional at a point x, and stores the result in y[0]
|
|
|
|
|
/// F.GetGradient() returns an operator that evaluates the gradient
|
|
|
|
|
/// F.GetGradient().GetGradient(x) returns an Hessian action operator.
|
|
|
|
|
///
|
|
|
|
|
/// The usual Operator::GetGradient(const Vector &x) method for this method
|
|
|
|
|
/// is deprecated as $ \nabla F $ only takes a single argument x.
|
|
|
|
|
/// It is redundant to use F.GetGradient(x).Mult(x, y) to evaluate the gradient.
|
|
|
|
|
/// Instead, use F.GetGradient().Mult(x, y) to evaluate the gradient at x.
|
|
|
|
|
///
|
|
|
|
|
/// The gradient and Hessian can be defined in two ways:
|
|
|
|
|
/// 1. If the gradient is available as a sperate operator,
|
|
|
|
|
/// then override Functional::GetGradient().
|
|
|
|
|
/// In this case, Functional::HessianMult() will not be called.
|
|
|
|
|
///
|
|
|
|
|
/// 2. Otherwise, override Functional::EvalGradient() and Functional::HessianMult()
|
|
|
|
|
/// to evaluate the gradient and Hessian action, respectively.
|
|
|
|
|
/// The helper classes, GradientOperator and HessianActionOperator,
|
|
|
|
|
/// will call these methods to evaluate the gradient and Hessian action.
|
|
|
|
|
/// If Hessian is a seperate operator, then you can override
|
|
|
|
|
/// The GradientOperator::GetGradient(x) will call Functional::GetHessian(x)
|
|
|
|
|
///
|
|
|
|
|
class Functional : public Operator
|
|
|
|
|
{
|
|
|
|
|
Operator * riesz_map = nullptr; ///< Riesz map operator, if available
|
|
|
|
|
public:
|
|
|
|
|
/// @brief Create a Functional with optional gradient and hessian
|
|
|
|
|
/// @param n number of variables
|
|
|
|
|
Functional(int n=0)
|
|
|
|
|
: Operator(1, n)
|
|
|
|
|
, grad_operator(*this)
|
|
|
|
|
, hessian_action_operator(*this)
|
|
|
|
|
{ }
|
|
|
|
|
|
|
|
|
|
#ifdef MFEM_USE_MPI
|
|
|
|
|
Functional(MPI_Comm comm, int n=0)
|
|
|
|
|
: Functional(n)
|
|
|
|
|
{ SetComm(comm); }
|
|
|
|
|
|
|
|
|
|
void SetComm(MPI_Comm comm_)
|
|
|
|
|
{
|
|
|
|
|
parallel = comm_ != MPI_COMM_NULL;
|
|
|
|
|
comm = comm_;
|
|
|
|
|
}
|
|
|
|
|
MPI_Comm GetComm() const { return comm; }
|
|
|
|
|
bool IsParallel() const { return parallel; }
|
|
|
|
|
#else
|
|
|
|
|
constexpr bool IsParallel() const { return false; }
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
void SetRieszMap(Operator &op) { riesz_map = &op; }
|
|
|
|
|
/// @brief return the GradientOperator that evaluates the gradient
|
|
|
|
|
/// input x is not used. Use GetGradient().Mult(x,y) to evaluate the gradient
|
|
|
|
|
/// we recommend using GetGradient() instead of GetGradient(x)
|
|
|
|
|
/// Deprecated. See Functional::GetGradient()
|
|
|
|
|
MFEM_DEPRECATED
|
|
|
|
|
Operator &GetGradient(const Vector &dummy) const override final { return GetGradient(); }
|
|
|
|
|
|
|
|
|
|
/// @brief Return the GradientOperator that wraps Functional::EvalGradient() for Mult().
|
|
|
|
|
/// @note If the functional has a corresponding standalone gradient operator,
|
|
|
|
|
/// override this method to return the gradient operator.
|
|
|
|
|
virtual Operator &GetGradient() const { return grad_operator; }
|
|
|
|
|
|
|
|
|
|
/// @brief Evaluate the functional at a point x that will be called by GradientOperator::Mult()
|
|
|
|
|
/// @note This method is not meant to be called directly. See, GradientOperator
|
|
|
|
|
virtual void EvalGradient(const Vector &x, Vector &y) const
|
|
|
|
|
{
|
|
|
|
|
MFEM_ABORT("Functional::EvalGradient() not implemented");
|
|
|
|
|
}
|
|
|
|
|
/// @brief Evaluate the Hessian action at a point x and direction d
|
|
|
|
|
/// that will be called by Functional::GetGradient().GetHessian(x).Mult(d,y)
|
|
|
|
|
/// @note This method is not meant to be called directly. See, HessianActionOperator
|
|
|
|
|
virtual void HessianMult(const Vector &x, const Vector &d, Vector &y) const
|
|
|
|
|
{
|
|
|
|
|
MFEM_ABORT("Functional::HessianMult() not implemented.");
|
|
|
|
|
}
|
|
|
|
|
/// @brief Return the HessianActionOperator at evaluation point x
|
|
|
|
|
/// that wraps Functional::HessianMult() for Mult().
|
|
|
|
|
/// See, HessianActionOperator and Functional::HessianMult().
|
|
|
|
|
///
|
|
|
|
|
/// @note If the Hessian is available as a seperate operator, override this method.
|
|
|
|
|
///
|
|
|
|
|
/// @warning If GetGradient() is overridden, this method will not be used.
|
|
|
|
|
virtual Operator &GetHessian(const Vector &x) const
|
|
|
|
|
{
|
|
|
|
|
hessian_action_operator.SetX(x);
|
|
|
|
|
return hessian_action_operator;
|
|
|
|
|
}
|
|
|
|
|
private:
|
|
|
|
|
#ifdef MFEM_USE_MPI
|
|
|
|
|
bool parallel=false;
|
|
|
|
|
#else
|
|
|
|
|
const static bool parallel=false;
|
|
|
|
|
#endif
|
|
|
|
|
#ifdef MFEM_USE_MPI
|
|
|
|
|
MPI_Comm comm;
|
|
|
|
|
#endif
|
|
|
|
|
/// @brief A helper class to return an operator that evaluates the gradient
|
|
|
|
|
/// using Functional::EvalGradient() method.
|
|
|
|
|
class GradientOperator : public Operator
|
|
|
|
|
{
|
|
|
|
|
private: const Functional &f; mutable Vector der;
|
|
|
|
|
public:
|
|
|
|
|
GradientOperator(const Functional &f) : Operator(f.Width()), f(f) {}
|
|
|
|
|
/// @brief Evaluate the gradient of Functional at a point x
|
|
|
|
|
void Mult(const Vector &x, Vector &y) const override final
|
|
|
|
|
{
|
|
|
|
|
if (f.riesz_map)
|
|
|
|
|
{
|
|
|
|
|
der.SetSize(f.Width());
|
|
|
|
|
f.EvalGradient(x, der);
|
|
|
|
|
f.riesz_map->Mult(der, y);
|
|
|
|
|
}
|
|
|
|
|
else
|
|
|
|
|
{
|
|
|
|
|
f.EvalGradient(x, y);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
/// @brief Evaluate the Hessian of Functional at a point x
|
|
|
|
|
Operator &GetGradient(const Vector &x) const override final { return f.GetHessian(x); }
|
|
|
|
|
};
|
|
|
|
|
friend class GradientOperator;
|
|
|
|
|
|
|
|
|
|
/// @brief A helper class to return an operator that applies the Hessian action
|
|
|
|
|
/// using Functional::HessianMult() method.
|
|
|
|
|
class HessianActionOperator : public Operator
|
|
|
|
|
{
|
|
|
|
|
private:
|
|
|
|
|
const Functional &f;
|
|
|
|
|
const Vector *x;
|
|
|
|
|
public:
|
|
|
|
|
HessianActionOperator(const Functional &f) : Operator(f.Width()), f(f) {}
|
|
|
|
|
void SetX(const Vector &new_x) { x = &new_x; }
|
|
|
|
|
void Mult(const Vector &d, Vector &y) const override { f.HessianMult(*x, d, y); }
|
|
|
|
|
};
|
|
|
|
|
friend class HessianActionOperator;
|
|
|
|
|
|
|
|
|
|
mutable GradientOperator grad_operator;
|
|
|
|
|
mutable HessianActionOperator hessian_action_operator;
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
/// @brief Stacked functioanl operator, [f1, ..., fk] where fi:R^n->R are functionals
|
|
|
|
|
/*
|
|
|
|
|
Typical usage of this class is to provide a single operator for multiple constraints.
|
|
|
|
|
For example, consider a minimization problem with k constraints,
|
|
|
|
|
min f0(u) s.t. fi(u)=0, i=1,...,k.
|
|
|
|
|
The Lagrangian functional is
|
|
|
|
|
$ L(u, lambda) = F0(u) + sum_i lambda_i * fi(u) $
|
|
|
|
|
The first-order optimality conditions are
|
|
|
|
|
$ \nabla f0(u) + \sum_i lambda_i * grad fi(u) = 0 $
|
|
|
|
|
$ fi(u) = 0 $
|
|
|
|
|
where lambda_i are the Lagrange multipliers.
|
|
|
|
|
The StackedFunctional class can be used to represent the list of constraints fi(u).
|
|
|
|
|
|
|
|
|
|
StackedFunctional::Mult(u, y) will evaluate each functional y[i]=fi(u)
|
|
|
|
|
|
|
|
|
|
StackedFunctional::GetGradient(u) represents an operator, column-stacked gradient
|
|
|
|
|
That is, [grad f0(u), ..., grad fk(u)] in R^{n x k}
|
|
|
|
|
If you want to extract the gradient as a matrix, use
|
|
|
|
|
StackedFunctional::GetGradientMatrix(const Vector &x, DenseMatrix &grad)
|
|
|
|
|
As functionals are not assumed to return a sparse vector, the gradient is dense.
|
|
|
|
|
|
|
|
|
|
StackedFunctional::GetGradient(u).Mult(lambda, y) contract the gradients with the Lagrange multipliers
|
|
|
|
|
y = sum lambda_i * grad fi(u)
|
|
|
|
|
StackedFunctional::GetGradient(u).MultTranspose(d, y) return the directional derivative for each k
|
|
|
|
|
y[i] = <grad fi(u), d>
|
|
|
|
|
|
|
|
|
|
StackedFunctional::GetHessian(u, lambda).Mult(d, y) will return the contracted Hessian action
|
|
|
|
|
$ y = \sum_i \lambda_i * H_{fi}(u, d) $
|
|
|
|
|
*/
|
|
|
|
|
/// @warning Functionals should be all serial or all parallel.
|
|
|
|
|
///
|
|
|
|
|
class StackedFunctional : public Operator
|
|
|
|
|
{
|
|
|
|
|
public:
|
|
|
|
|
StackedFunctional(int n=0)
|
|
|
|
|
: Operator(0, n)
|
|
|
|
|
, funcs(0)
|
|
|
|
|
, grad_helper_op(*this)
|
|
|
|
|
, hessian_helper_op(*this)
|
|
|
|
|
{}
|
|
|
|
|
StackedFunctional(Functional &f)
|
|
|
|
|
: Operator(0, f.Width())
|
|
|
|
|
, grad_helper_op(*this)
|
|
|
|
|
, hessian_helper_op(*this)
|
|
|
|
|
{ AddFunctional(f); }
|
|
|
|
|
|
|
|
|
|
StackedFunctional(const std::vector<Functional*> &funcs)
|
|
|
|
|
: Operator((int)funcs.size(), funcs[0]->Width())
|
|
|
|
|
, grad_helper_op(*this)
|
|
|
|
|
, hessian_helper_op(*this)
|
|
|
|
|
{ for (auto &f : funcs) { AddFunctional(*f); } }
|
|
|
|
|
|
|
|
|
|
void AddFunctional(Functional &f)
|
|
|
|
|
{
|
|
|
|
|
#ifdef MFEM_USE_MPI
|
|
|
|
|
if (funcs.empty()) { if (f.IsParallel()) { SetComm(f.GetComm()); } }
|
|
|
|
|
#endif
|
|
|
|
|
MFEM_VERIFY(f.Width() == Width(),
|
|
|
|
|
"StackedFunctional::AddFunctional: Functional width does not match with the operator.");
|
|
|
|
|
MFEM_VERIFY(parallel == f.IsParallel(),
|
|
|
|
|
"StackedFunctional::AddFunctional: Parallelism mismatch.");
|
|
|
|
|
funcs.push_back(&f);
|
|
|
|
|
height++;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void Mult(const Vector &x, Vector &y) const override
|
|
|
|
|
{
|
|
|
|
|
y.SetSize(Height());
|
|
|
|
|
Vector yview;
|
|
|
|
|
for (int i=0; i<Height(); i++)
|
|
|
|
|
{
|
|
|
|
|
yview.MakeRef(y, i, 1);
|
|
|
|
|
funcs[i]->Mult(x, yview);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
Operator &GetGradient(const Vector &x) const override
|
|
|
|
|
{
|
|
|
|
|
grad_helper_op.SetX(x);
|
|
|
|
|
return grad_helper_op;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void GetGradientMatrix(const Vector &x, DenseMatrix &grads) const
|
|
|
|
|
{
|
|
|
|
|
grads.SetSize(Width(), Height());
|
|
|
|
|
Vector grad;
|
|
|
|
|
for (int i=0; i<Height(); i++)
|
|
|
|
|
{
|
|
|
|
|
grads.GetColumnReference(i, grad);
|
|
|
|
|
funcs[i]->GetGradient().Mult(x, grad);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
Functional &GetFunctional(int i) const
|
|
|
|
|
{
|
|
|
|
|
MFEM_VERIFY(i >= 0 && i < Height(),
|
|
|
|
|
"StackedFunctional::GetFunctional: Index out of bounds.");
|
|
|
|
|
return *funcs[i];
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
Operator &GetHessian(const Vector &x, const Vector &lambda) const
|
|
|
|
|
{
|
|
|
|
|
hessian_helper_op.SetX(x, lambda);
|
|
|
|
|
return hessian_helper_op;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
bool parallel;
|
|
|
|
|
bool IsParallel() const { return parallel; }
|
|
|
|
|
#ifdef MFEM_USE_MPI
|
|
|
|
|
void SetComm(MPI_Comm comm_)
|
|
|
|
|
{
|
|
|
|
|
parallel = comm != MPI_COMM_NULL;
|
|
|
|
|
comm = comm_;
|
|
|
|
|
}
|
|
|
|
|
MPI_Comm GetComm() const { return comm; }
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
protected:
|
|
|
|
|
#ifdef MFEM_USE_MPI
|
|
|
|
|
MPI_Comm comm;
|
|
|
|
|
#endif
|
|
|
|
|
std::vector<Functional*> funcs;
|
|
|
|
|
|
|
|
|
|
class GradientOperator : public Operator
|
|
|
|
|
{
|
|
|
|
|
public:
|
|
|
|
|
GradientOperator(const StackedFunctional &op)
|
|
|
|
|
: Operator(op.Width(), op.Height())
|
|
|
|
|
, op(op)
|
|
|
|
|
, tmp_grad(op.Width())
|
|
|
|
|
{}
|
|
|
|
|
void SetX(const Vector &x) const { x_curr = &x; }
|
|
|
|
|
Operator &GetGradient(const Vector &lambda) const override
|
|
|
|
|
{
|
|
|
|
|
op.hessian_helper_op.SetX(*x_curr, lambda);
|
|
|
|
|
return op.hessian_helper_op;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void Mult(const Vector &lambda, Vector &y) const override
|
|
|
|
|
{
|
|
|
|
|
y.SetSize(op.Width());
|
|
|
|
|
y = 0.0;
|
|
|
|
|
for (int i=0; i<op.Height(); i++)
|
|
|
|
|
{
|
|
|
|
|
op.funcs[i]->GetGradient().Mult(*x_curr, tmp_grad);
|
|
|
|
|
y.Add(lambda[i], tmp_grad);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
void MultTranspose(const Vector &x, Vector &y) const override
|
|
|
|
|
{
|
|
|
|
|
y.SetSize(op.Height());
|
|
|
|
|
for (int i=0; i<op.Height(); i++)
|
|
|
|
|
{
|
|
|
|
|
op.funcs[i]->GetGradient().Mult(x, tmp_grad);
|
|
|
|
|
y[i] = InnerProduct(tmp_grad, *x_curr);
|
|
|
|
|
}
|
|
|
|
|
#ifdef MFEM_USE_MPI
|
|
|
|
|
if (op.IsParallel())
|
|
|
|
|
{
|
|
|
|
|
MPI_Allreduce(MPI_IN_PLACE, y.GetData(), op.Height(),
|
|
|
|
|
MPITypeMap<real_t>::mpi_type, MPI_SUM,
|
|
|
|
|
op.GetComm());
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
private:
|
|
|
|
|
const StackedFunctional &op;
|
|
|
|
|
mutable const Vector *x_curr;
|
|
|
|
|
mutable Vector tmp_grad;
|
|
|
|
|
|
|
|
|
|
};
|
|
|
|
|
class HessianActionOperator : public Operator
|
|
|
|
|
{
|
|
|
|
|
public:
|
|
|
|
|
HessianActionOperator(const StackedFunctional &op)
|
|
|
|
|
: Operator(op.Width()), op(op)
|
|
|
|
|
{}
|
|
|
|
|
void SetX(const Vector &x, const Vector &lambda) const { x_curr = &x; lambda_curr = λ }
|
|
|
|
|
|
|
|
|
|
void Mult(const Vector &d, Vector &y) const override
|
|
|
|
|
{
|
|
|
|
|
y.SetSize(op.Width());
|
|
|
|
|
y = 0.0;
|
|
|
|
|
for (int i=0; i<op.Height(); i++)
|
|
|
|
|
{
|
|
|
|
|
op.funcs[i]->GetGradient().GetGradient(*x_curr).Mult(d, tmp_hessian);
|
|
|
|
|
y.Add((*lambda_curr)[i], tmp_hessian);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
private:
|
|
|
|
|
const StackedFunctional &op;
|
|
|
|
|
mutable Vector tmp_hessian;
|
|
|
|
|
mutable const Vector *x_curr;
|
|
|
|
|
mutable const Vector *lambda_curr;
|
|
|
|
|
};
|
|
|
|
|
friend class GradientOperator;
|
|
|
|
|
friend class HessianActionOperator;
|
|
|
|
|
mutable GradientOperator grad_helper_op;
|
|
|
|
|
mutable HessianActionOperator hessian_helper_op;
|
|
|
|
|
private:
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
class ConstrainedOptimizationProblem : public Functional
|
|
|
|
|
{
|
|
|
|
|
public:
|
|
|
|
|
ConstrainedOptimizationProblem(Functional &objective_,
|
|
|
|
|
Operator *eq_constraints_=nullptr,
|
|
|
|
|
Operator *ineq_constraints_=nullptr)
|
|
|
|
|
: Functional(objective_.Width())
|
|
|
|
|
, objective(objective_)
|
|
|
|
|
, eq_constraints(eq_constraints_)
|
|
|
|
|
, ineq_constraints(ineq_constraints_)
|
|
|
|
|
{
|
|
|
|
|
// Check Size
|
|
|
|
|
MFEM_VERIFY((eq_constraints == nullptr ||
|
|
|
|
|
eq_constraints->Width() == objective.Width()),
|
|
|
|
|
"ConstrainedFunctional: Equality constraints width does not match with the objective.");
|
|
|
|
|
MFEM_VERIFY((ineq_constraints == nullptr ||
|
|
|
|
|
ineq_constraints->Width() == objective.Width()),
|
|
|
|
|
"ConstrainedFunctional: Inequality constraints width does not match with the objective.");
|
|
|
|
|
#ifdef MFEM_USE_MPI
|
|
|
|
|
if (objective.IsParallel()) { SetComm(objective.GetComm()); }
|
|
|
|
|
#endif
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
Functional &GetObjective() { return objective; }
|
|
|
|
|
const Functional &GetObjective() const { return objective; }
|
|
|
|
|
Operator *GetEqualityConstraints() { return eq_constraints; }
|
|
|
|
|
const Operator *GetEqualityConstraints() const { return eq_constraints; }
|
|
|
|
|
Operator *GetInequalityConstraints() { return ineq_constraints; }
|
|
|
|
|
const Operator *GetInequalityConstraints() const { return ineq_constraints; }
|
|
|
|
|
protected:
|
|
|
|
|
Functional &objective;
|
|
|
|
|
Operator *eq_constraints;
|
|
|
|
|
Operator *ineq_constraints;
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
/// @brief A Lagrangian functional for
|
|
|
|
|
/// min F(u)
|
|
|
|
|
/// subject to C(u) = 0
|
|
|
|
|
/// That is, L(u, lambda) = F(u) + <lambda, C(u)>
|
|
|
|
|
///
|
|
|
|
|
/// We assume that $ F:R^n -> R $ is a functional,
|
|
|
|
|
/// $ C:R^n -> R^k $ is an equality constraint operator,
|
|
|
|
|
/// C should return a residual. That is,
|
|
|
|
|
/// C(u) = c, then C.Mult(u, y) should return y[i] = C_i(u) - c_i.
|
|
|
|
|
///
|
|
|
|
|
/// C.GetGradient(u):R^k -> R^n that takes lambda and returns the contracted gradient at x
|
|
|
|
|
/// C.GetGradient(u).Mult(lambda, y) returns y = sum lambda_i * grad C_i(u)
|
|
|
|
|
///
|
|
|
|
|
/// C's gradient should support MultTranspose method
|
|
|
|
|
/// That is, C.GetGradient(u).MultTranspose(d, y) returns y[i] = <grad C_i(u), d>
|
|
|
|
|
///
|
|
|
|
|
class LagrangianFunctional : public ConstrainedOptimizationProblem
|
|
|
|
|
{
|
|
|
|
|
private:
|
|
|
|
|
mutable Vector eq_residual;
|
|
|
|
|
public:
|
|
|
|
|
LagrangianFunctional(Functional &objective,
|
|
|
|
|
Operator &eq_constraints)
|
|
|
|
|
: ConstrainedOptimizationProblem(objective, &eq_constraints)
|
|
|
|
|
, eq_residual(eq_constraints.Height())
|
|
|
|
|
{
|
|
|
|
|
width = objective.Width() + eq_constraints.Height();
|
|
|
|
|
|
|
|
|
|
offsets.SetSize(3);
|
|
|
|
|
offsets[0] = 0;
|
|
|
|
|
offsets[1] = objective.Width();
|
|
|
|
|
offsets[2] = eq_constraints.Height();
|
|
|
|
|
offsets.PartialSum();
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void Mult(const Vector &x, Vector &y) const override
|
|
|
|
|
{
|
|
|
|
|
const BlockVector input_block(const_cast<Vector&>(x), offsets);
|
|
|
|
|
const Vector &u = input_block.GetBlock(0);
|
|
|
|
|
const Vector &lambda = input_block.GetBlock(1);
|
|
|
|
|
|
|
|
|
|
y.SetSize(1);
|
|
|
|
|
y[0] = 0.0;
|
|
|
|
|
objective.Mult(u, y);
|
|
|
|
|
eq_constraints->Mult(u, eq_residual);
|
|
|
|
|
y[0] += InnerProduct(lambda, eq_residual);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void EvalGradient(const Vector &x, Vector &y) const override
|
|
|
|
|
{
|
|
|
|
|
const BlockVector input_block(const_cast<Vector&>(x), offsets);
|
|
|
|
|
const Vector &u = input_block.GetBlock(0);
|
|
|
|
|
const Vector &lambda = input_block.GetBlock(1);
|
|
|
|
|
|
|
|
|
|
y.SetSize(Width());
|
|
|
|
|
BlockVector output_block(y, offsets);
|
|
|
|
|
Vector &opt_residual = output_block.GetBlock(0);
|
|
|
|
|
eq_residual = output_block.GetBlock(1);
|
|
|
|
|
y = 0.0;
|
|
|
|
|
// grad F(u) + \sum_i lambda_i grad C_i(u)
|
|
|
|
|
objective.GetGradient().Mult(u, opt_residual);
|
|
|
|
|
eq_constraints->GetGradient(u).AddMult(lambda, opt_residual);
|
|
|
|
|
// grad C_i(u)^T
|
|
|
|
|
eq_constraints->GetGradient(u).MultTranspose(u, eq_residual);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/// @brief Evaluate the Hessian action at a point x=[u, lambda]
|
|
|
|
|
/// and direction d=[v, mu]
|
|
|
|
|
/// $ [H_F(u,d) + \sum_i lambda_i H_{C_i}(u, d) + <\nabla C_i(u), mu> $
|
|
|
|
|
void HessianMult(const Vector &x, const Vector &d, Vector &y) const override
|
|
|
|
|
{
|
|
|
|
|
const BlockVector input_block(const_cast<Vector&>(x), offsets);
|
|
|
|
|
const Vector &u = input_block.GetBlock(0);
|
|
|
|
|
const Vector &lambda = input_block.GetBlock(1);
|
|
|
|
|
|
|
|
|
|
const BlockVector direction_block(const_cast<Vector&>(x), offsets);
|
|
|
|
|
const Vector &v = direction_block.GetBlock(0);
|
|
|
|
|
const Vector &mu = direction_block.GetBlock(1);
|
|
|
|
|
|
|
|
|
|
y.SetSize(Width());
|
|
|
|
|
BlockVector output_block(y, offsets);
|
|
|
|
|
Vector &opt_H = output_block.GetBlock(0); // Optimality Hessian
|
|
|
|
|
Vector &eq_H = output_block.GetBlock(1); // Equality Hessian
|
|
|
|
|
|
|
|
|
|
// H_F(u,d) + \sum_i lambda_i H_{C_i}(u, d) + mu_i grad C_i(u)
|
|
|
|
|
objective.GetGradient().GetGradient(u).Mult(v, opt_H);
|
|
|
|
|
eq_constraints->GetGradient(u).GetGradient(lambda).AddMult(v, opt_H);
|
|
|
|
|
eq_constraints->GetGradient(u).Mult(mu, eq_H);
|
|
|
|
|
// <grad C_i(u), d>
|
|
|
|
|
eq_constraints->GetGradient(u).MultTranspose(d, eq_H);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
protected:
|
|
|
|
|
Array<int> offsets; // offsets for [x, lambda, mu]
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
/// @brief An augmented Lagrangian functional of the form
|
|
|
|
|
/// F(u) + 0.5 mu * ||C(u)||^2 + <lambda, C(u)>
|
|
|
|
|
/// where F is the objective functional,
|
|
|
|
|
/// C is the equality constraint operator,
|
|
|
|
|
/// lambda is the Lagrange multiplier vector (initialized to zero),
|
|
|
|
|
/// mu is the penalty parameter (defaults to 1.0)
|
|
|
|
|
///
|
|
|
|
|
/// Currently, only equality constraints are supported.
|
|
|
|
|
///
|
|
|
|
|
/// AugLagrangianFunctional::Update() will update the penalty and Lagrange multiplier vectors
|
|
|
|
|
/// By default, lambda <- lambda + mu * C(u)
|
|
|
|
|
/// mu <- mu (no update)
|
|
|
|
|
class AugLagrangianFunctional : public ConstrainedOptimizationProblem
|
|
|
|
|
{
|
|
|
|
|
public:
|
|
|
|
|
AugLagrangianFunctional(Functional &objective_,
|
|
|
|
|
Operator &eq_constraints_)
|
|
|
|
|
: ConstrainedOptimizationProblem(objective_, &eq_constraints_)
|
|
|
|
|
, lambda(eq_constraints_.Height())
|
|
|
|
|
, mu(1.0)
|
|
|
|
|
, eq_residual(eq_constraints_.Height())
|
|
|
|
|
, eq_dir(eq_constraints_.Height())
|
|
|
|
|
{
|
|
|
|
|
lambda = 0.0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void SetLambda(const Vector &lambda_)
|
|
|
|
|
{
|
|
|
|
|
MFEM_VERIFY(lambda_.Size() == eq_constraints->Height(),
|
|
|
|
|
"AugLagrangianFunctional: Lambda size does not match with the equality constraints.");
|
|
|
|
|
lambda = lambda_;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void SetPenalty(real_t mu_)
|
|
|
|
|
{
|
|
|
|
|
MFEM_VERIFY(mu_ >= 0.0,
|
|
|
|
|
"AugLagrangianFunctional: Penalty parameter mu must be non-negative.");
|
|
|
|
|
mu = mu_;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
virtual void Update(const Vector &x)
|
|
|
|
|
{
|
|
|
|
|
// Update the Lagrange multipliers
|
|
|
|
|
eq_constraints->AddMult(x, lambda, mu);
|
|
|
|
|
// Update the penalty parameter
|
|
|
|
|
// Do nothing
|
|
|
|
|
}
|
|
|
|
|
const Vector &GetLambda() const { return lambda; }
|
|
|
|
|
real_t GetPenalty() const { return mu; }
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
void Mult(const Vector &x, Vector &y) const override
|
|
|
|
|
{
|
|
|
|
|
y.SetSize(1);
|
|
|
|
|
objective.Mult(x, y);
|
|
|
|
|
eq_constraints->Mult(x, eq_residual);
|
|
|
|
|
y[0] += lambda*eq_residual;
|
|
|
|
|
y[0] += 0.5 * mu * (eq_residual*eq_residual);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void EvalGradient(const Vector &x, Vector &y) const override
|
|
|
|
|
{
|
|
|
|
|
y.SetSize(Width());
|
|
|
|
|
// grad F(x) + \sum_i (lambda_i + mu * C_i(x)) grad C_i(x)
|
|
|
|
|
Vector curr_lambda = lambda; // store lambda + mu * C(x)
|
|
|
|
|
objective.GetGradient().Mult(x, y);
|
|
|
|
|
eq_constraints->Mult(x, eq_residual);
|
|
|
|
|
curr_lambda.Add(mu, eq_residual);
|
|
|
|
|
eq_constraints->GetGradient(x).AddMult(curr_lambda, y);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/// @brief Evaluate the Hessian action at a point x=[u, lambda]
|
|
|
|
|
/// and direction d=[v, mu]
|
|
|
|
|
/// $ H_F(u,d) + \sum_i \lambda_i H_{C_i}(u, d) + <\nabla C_i(u), mu> $
|
|
|
|
|
void HessianMult(const Vector &x, const Vector &d, Vector &y) const override
|
|
|
|
|
{
|
|
|
|
|
// H_F(u,d) + \sum_i lambda_i H_{C_i}(u, d) + mu_i grad C_i(u) <grad C_i(u), d>
|
|
|
|
|
objective.GetGradient().GetGradient(x).Mult(d, y);
|
|
|
|
|
|
|
|
|
|
Vector curr_lambda = lambda;
|
|
|
|
|
eq_constraints->Mult(x, eq_residual);
|
|
|
|
|
curr_lambda.Add(mu, eq_residual);
|
|
|
|
|
|
|
|
|
|
eq_constraints->GetGradient(x).GetGradient(curr_lambda).AddMult(d, y);
|
|
|
|
|
// eq_dir = <grad C_i(u), d>
|
|
|
|
|
eq_constraints->GetGradient(x).MultTranspose(d, eq_dir);
|
|
|
|
|
// mu_i <grad C_i(u), eq_dir>
|
|
|
|
|
eq_constraints->GetGradient(x).AddMult(eq_dir, y, mu);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
protected:
|
|
|
|
|
Vector lambda;
|
|
|
|
|
real_t mu;
|
|
|
|
|
mutable Vector eq_residual; // residual of the equality constraints, R^k
|
|
|
|
|
// directional derivative of the equality constraints, R^k
|
|
|
|
|
mutable Vector eq_dir;
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
/// @brief Quadratic functional of the form
|
|
|
|
|
/// f(u) = 0.5 * <A u, u> + beta<b, u> + c
|
|
|
|
|
/// where A is a square (possibly nonlinear) operator,
|
|
|
|
|
/// beta is a scalar (defaults to 1.0, not used when b is nullptr),
|
|
|
|
|
/// b is a vector (independent of u, optional),
|
|
|
|
|
/// c is a constant (independent of u, optional).
|
|
|
|
|
/// GetHessian() returns the operator A.
|
|
|
|
|
///
|
|
|
|
|
class QuadraticFunctional : public Functional
|
|
|
|
|
{
|
|
|
|
|
public:
|
|
|
|
|
QuadraticFunctional()
|
|
|
|
|
: Functional(0)
|
|
|
|
|
, A(nullptr), b(nullptr), c(0.0)
|
|
|
|
|
{}
|
|
|
|
|
QuadraticFunctional(const Operator *A_, const Vector *b_=nullptr,
|
|
|
|
|
const real_t beta_=1.0,
|
|
|
|
|
const real_t c_=0.0);
|
|
|
|
|
#ifdef MFEM_USE_MPI
|
|
|
|
|
QuadraticFunctional(MPI_Comm comm_)
|
|
|
|
|
: QuadraticFunctional()
|
|
|
|
|
{ SetComm(comm_); }
|
|
|
|
|
|
|
|
|
|
QuadraticFunctional(MPI_Comm comm_, const Operator *A_,
|
|
|
|
|
const Vector *b_=nullptr,
|
|
|
|
|
const real_t beta_=1.0, const real_t c_=0.0)
|
|
|
|
|
: QuadraticFunctional(A_, b_, beta_, c_)
|
|
|
|
|
{ SetComm(comm_); }
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
void SetOperator(const Operator &A_);
|
|
|
|
|
void SetVector(const Vector &b_, const real_t beta_=1.0);
|
|
|
|
|
void SetConstant(real_t c_);
|
|
|
|
|
|
|
|
|
|
void Mult(const Vector &x, Vector &y) const override;
|
|
|
|
|
void EvalGradient(const Vector &x, Vector &y) const override;
|
|
|
|
|
protected:
|
|
|
|
|
const Operator *A;
|
|
|
|
|
real_t beta;
|
|
|
|
|
const Vector *b;
|
|
|
|
|
real_t c;
|
|
|
|
|
mutable Vector aux;
|
|
|
|
|
protected:
|
|
|
|
|
/// @brief return the underlying Operator A
|
|
|
|
|
/// @warning Modifying the returned operator leads to undefined behavior.
|
|
|
|
|
Operator& GetHessian(const Vector &dummy) const override
|
|
|
|
|
{
|
|
|
|
|
return const_cast<Operator&>(*A);
|
|
|
|
|
}
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
class Optimizer : public IterativeSolver
|
|
|
|
|
{
|
|
|
|
|
public:
|
|
|
|
|
Optimizer() : IterativeSolver(), f(nullptr) { }
|
|
|
|
|
#ifdef MFEM_USE_MPI
|
|
|
|
|
Optimizer(MPI_Comm comm) : IterativeSolver(comm), f(nullptr) { }
|
|
|
|
|
#endif
|
|
|
|
|
// @brief Set the subproblem functional operator
|
|
|
|
|
// @param op the functional operator
|
|
|
|
|
// @note The functional will be stored in subproblem, and oper will be set to the gradient of the functional.
|
|
|
|
|
void SetOperator(const Functional &f_)
|
|
|
|
|
{
|
|
|
|
|
f = &f_;
|
|
|
|
|
IterativeSolver::SetOperator(f_.GetGradient());
|
|
|
|
|
}
|
|
|
|
|
virtual void SetLinearSolver(Solver &prec) { IterativeSolver::SetPreconditioner(prec); }
|
|
|
|
|
|
|
|
|
|
/// @brief This will abort. Should be called only with a Functional operator.
|
|
|
|
|
void SetOperator(const Operator &op) override
|
|
|
|
|
{
|
|
|
|
|
MFEM_ABORT("OptSolver::SetOperator() should not be called directly. Use SetFunctional() instead.");
|
|
|
|
|
}
|
|
|
|
|
protected:
|
|
|
|
|
const Functional * f;
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
class NewtonOptimizer : public Optimizer
|
|
|
|
|
{
|
|
|
|
|
private:
|
|
|
|
|
real_t step_size = 1.0; // default step size
|
|
|
|
|
public:
|
|
|
|
|
NewtonOptimizer() : Optimizer() { }
|
|
|
|
|
#ifdef MFEM_USE_MPI
|
|
|
|
|
NewtonOptimizer(MPI_Comm comm) : Optimizer(comm) { }
|
|
|
|
|
#endif
|
|
|
|
|
void SetStepSize(real_t step_size_) { step_size = step_size_; }
|
|
|
|
|
|
|
|
|
|
void Mult(const Vector &x, Vector &y) const override
|
|
|
|
|
{
|
|
|
|
|
dx.SetSize(x.Size());
|
|
|
|
|
y.SetSize(x.Size());
|
|
|
|
|
y = x;
|
|
|
|
|
MFEM_ASSERT(f != nullptr,
|
|
|
|
|
"NewtonOptimizer::Mult() called without a functional operator.");
|
|
|
|
|
MFEM_ASSERT(prec != nullptr,
|
|
|
|
|
"NewtonOptimizer::Mult() called without a linear solver.");
|
|
|
|
|
for (int i=0; i<max_iter; i++)
|
|
|
|
|
{
|
|
|
|
|
oper->Mult(y, grad);
|
|
|
|
|
Operator &hess = oper->GetGradient(y);
|
|
|
|
|
prec->SetOperator(hess);
|
|
|
|
|
prec->Mult(grad, dx);
|
|
|
|
|
y.Add(-step_size, dx);
|
|
|
|
|
if (Dot(dx, dx) < abs_tol*abs_tol)
|
|
|
|
|
{
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
private:
|
|
|
|
|
mutable Vector grad;
|
|
|
|
|
mutable Vector dx;
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
class GradientDescentOptimizer : public Optimizer
|
|
|
|
|
{
|
|
|
|
|
private:
|
|
|
|
|
real_t step_size = 1.0; // default step size
|
|
|
|
|
public:
|
|
|
|
|
GradientDescentOptimizer() : Optimizer() { }
|
|
|
|
|
#ifdef MFEM_USE_MPI
|
|
|
|
|
GradientDescentOptimizer(MPI_Comm comm) : Optimizer(comm) { }
|
|
|
|
|
#endif
|
|
|
|
|
void SetStepSize(real_t step_size_) { step_size = step_size_; }
|
|
|
|
|
|
|
|
|
|
void Mult(const Vector &x, Vector &y) const override
|
|
|
|
|
{
|
|
|
|
|
grad.SetSize(x.Size());
|
|
|
|
|
y.SetSize(x.Size());
|
|
|
|
|
y = x;
|
|
|
|
|
MFEM_ASSERT(f != nullptr,
|
|
|
|
|
"NewtonOptimizer::Mult() called without a functional operator.");
|
|
|
|
|
MFEM_ASSERT(prec != nullptr,
|
|
|
|
|
"NewtonOptimizer::Mult() called without a linear solver.");
|
|
|
|
|
for (int i=0; i<max_iter; i++)
|
|
|
|
|
{
|
|
|
|
|
oper->Mult(y, grad);
|
|
|
|
|
y.Add(-step_size, grad);
|
|
|
|
|
if (Dot(grad, grad) < abs_tol*abs_tol)
|
|
|
|
|
{
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
private:
|
|
|
|
|
mutable Vector grad;
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
} // namespace mfem
|
|
|
|
|
#endif // MFEM_FUNCTIONAL_HPP
|