apop_mle_settings Struct Reference

#include <apop.h>

Data Fields
double	delta
double	dim_cycle_tolerance
int	iters_fixed_T
double	k
int	max_iterations
char *	method
double	mu_t
int	n_tries
apop_data **	path
gsl_rng *	rng
double *	starting_pt
double	step_size
double	t_initial
double	t_min
double	tolerance
int	verbose

Detailed Description

The settings for maximum likelihood estimation (including simulated annealing).

Field Documentation

dim_cycle_tolerance

double apop_mle_settings::dim_cycle_tolerance

If zero (the default), the usual procedure. If , cycle across dimensions: fix all but the first dimension at the starting point, optimize only the first dim. Then fix the all but the second dim, and optimize the second dim. Continue through all dims, until the log likelihood at the outset of one cycle through the dimensions is within this amount of the previous cycle's log likelihood. There will be at least two cycles.

max_iterations

int apop_mle_settings::max_iterations

Ignored by simulated annealing. Other methods halt (and set the "status" element of the output estimate's info page) if they do this many iterations without finding an optimum.

method

char * apop_mle_settings::method

The method to be used for the optimization. All strings are case-insensitive.

String

<Name>

Notes

"NM simplex"

Nelder-Mead simplex

Does not use gradients at all. Can sometimes get stuck.

"FR cg"

Conjugate gradient (Fletcher-Reeves) (default)

CG methods use derivatives. The converge to the optimum of a quadratic function in one step; performance degrades as the objective digresses from quadratic.

"BFGS cg"

Broyden-Fletcher-Goldfarb-Shanno conjugate gradient

"PR cg"

Polak-Ribiere conjugate gradient

"Annealing"

simulated annealing

Slow but works for objectives of arbitrary complexity, including stochastic objectives.

"Newton"

Newton's method

Search by finding a root of the derivative. Expects that gradient is reasonably well-behaved.

"Newton hybrid"

Newton's method/gradient descent hybrid

Find a root of the derivative via the Hybrid method

If Newton proposes stepping outside of a certain interval, use an alternate method. See the GSL manual for discussion.

"Newton hybrid no scale"

Newton's method/gradient descent hybrid with spherical scale

As above, but use a simplified trust region.

path

apop_data ** apop_mle_settings::path

If not NULL, record each vector tried by the optimizer as one row of this apop_data set. Each row of the matrix element holds the vector tried; the corresponding element in the vector is the evaluated value at that vector (after out-of-constraints penalties have been subtracted). A new apop_data set is allocated at the pointer you send in. This data set has no names; add them as desired. For a sample use, see Optimization.

starting_pt

double * apop_mle_settings::starting_pt

An array of doubles (e.g., (double*){2,4,6,8}) suggesting a starting point. If NULL, use an all-ones vector. If startv is a gsl_vector and is not a view of a matrix, use .starting_pt=startv->data.

step_size

double apop_mle_settings::step_size

The initial step size.

tolerance

double apop_mle_settings::tolerance

The precision the minimizer uses in its stopping rule. Only vaguely related to the precision of the actual MLE.

verbose

int apop_mle_settings::verbose

Give status updates as we go. This is orthogonal to the apop_opts.verbose setting.