Optimization::ConstrainedMinimizationProblem Struct Reference

Same as MinimizationProblem but with nonlinear constraints C(x)=0 and D(x)<=0, bound constraints bmin/bmax. More...

#include <Minimization.h>

Public Member Functions
	ConstrainedMinimizationProblem (ScalarFieldFunction *f, VectorFieldFunction *C, VectorFieldFunction *D)
ConvergenceResult	SolveGD (int &maxIters)
	Gradient descent.
ConvergenceResult	StepGD ()
	Returns MaxItersReached on normal exit.
ConvergenceResult	SolveNewton (int &maxIters)
	Newton iteration.
ConvergenceResult	StepNewton ()
	Returns MaxItersReached on normal exit.
ConvergenceResult	StepTR (Real R)
void	NullspaceProjection (const Vector &x, Vector &dx)
	Helper to project grad onto the constraint nullspace.
ConvergenceResult	LineMinimizationStep (const Vector &dx, Real &alpha0)
	Helper for line minimzation – like above, but on the manifold surface.
bool	CheckPoint (const Vector &x) const
	Helper to check point feasibility.
bool	SolveFeasiblePoint (Vector &x, int maxIters, ConvergenceResult *res=NULL)
	Helper to solve for a feasible point.

Public Attributes
ScalarFieldFunction *	f
VectorFieldFunction *	C
VectorFieldFunction *	D
Vector	bmin
Vector	bmax
Vector	x
Real	tolx
Real	tolf
Real	tolgrad
Real	tolc
bool	sparse
int	verbose
std::vector< Vector > *	S
Real	fx
Vector	grad
Matrix	H
ConstrainedNewtonRoot	rootSolver

Detailed Description

Same as MinimizationProblem but with nonlinear constraints C(x)=0 and D(x)<=0, bound constraints bmin/bmax.

Performs the line search while maintaining feasibility on the constraint manifold. The initial point must lie in the feasibility set. This isn't really a well established minimization technique and hasn't been tested thoroughly.

StepTR performs a sequential linear programming step with a given trust region. We could implement SQP, but other libraries would probably be better.

If sparse is set to true, C and D are assumed to be subclasses of SparseVectorFunction. Undefined behavior if this does not hold!

Member Function Documentation

ConvergenceResult Optimization::ConstrainedMinimizationProblem::StepTR

(

Real

R

)

Gradient-based trust-region method Returns MaxItersReached on normal exit, LocalMinimum if R should be increased, Divergence if it should be reduced

---

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

• Minimization.h