drop_convergence_test Class Template Reference

Convergence test based on monitoring an objective function. More...

#include <drop_convergence_test.h>

List of all members.

Public Member Functions
	drop_convergence_test (SizeType n_test_points=5, FloatType max_drop_eps=FloatType(1.e-5), FloatType iteration_coefficient=FloatType(2))
	Constructor.
SizeType	n_test_points () const
	Number of most recent objective function values used in fit of straight lines (as passed to the constructor).
FloatType	max_drop_eps () const
	Base tolerance for test of drop of objective function (as passed to the constructor).
FloatType	iteration_coefficient () const
	Tolerance for comparison of slopes (as passed to the constructor).
bool	operator() (FloatType f)
	Execution of the convergence test.
af::shared< FloatType >	objective_function_values () const
	Values of the objective function monitored so far.
FloatType	max_drop () const
	Maximum drop monitored so far.

---

Detailed Description

template<typename FloatType, typename SizeType = std::size_t>
class scitbx::lbfgs::drop_convergence_test< FloatType, SizeType >

Convergence test based on monitoring an objective function.

This test monitors an objective function f in the course of the minimization to determine when it has reached a plateau. The test is independent of the absolute scale of the objective function.

• The maximum value of the drop in the objective function is determined (max_drop()).

• A straight line is fitted to the last n_test_points() values of the objective function. Let slope be the slope of this fitted line. Convergence is detected if:

            -slope <=   max_drop() * max_drop_eps()
   number_of_iterations^iteration_coefficient()
            

  Note that with iteration_coefficient() > 1 this test becomes increasingly tolerant with the number of iterations. The rationale is that the slope will not change very much for parameters that are already close to the optimum.

---

Constructor & Destructor Documentation

drop_convergence_test	(	SizeType	n_test_points = 5,
		FloatType	max_drop_eps = FloatType(1.e-5),
		FloatType	iteration_coefficient = FloatType(2)
	)			[inline, explicit]

Constructor.

See the class details for an outline of the convergence detection algorithm and the meaning of the parameters.

References drop_convergence_test::n_test_points(), and SCITBX_ASSERT.

---

Member Function Documentation

bool operator()

(

FloatType

f

)

[inline]

Execution of the convergence test.

Note that at least n_test_points() executions are required before convergence will be evaluated.

References SCITBX_ASSERT.

---

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

• scitbx/lbfgs/drop_convergence_test.h