Plugin type: minimizer/singlecost

Plugin type: minimizer/singlecost
Prev	Chapter 3. Plugin Reference

These plug-ins provide optimizers of many-to-one functions

Plugins:

gdas

Gradient descent with automatic step size correction.. Supported parameters are:

Name	Type	Default	Description
ftolr	double in [0, inf)	0	Stop if the relative change of the criterion is below.
max-step	double in (0, inf)	2	Maximal absolute step size
maxiter	uint in [1, inf)	200	Stopping criterion: the maximum number of iterations
min-step	double in (0, inf)	0.1	Minimal absolute step size
xtola	double in [0, inf)	0.01	Stop if the inf-norm of the change applied to x is below this value.

gdsq

Gradient descent with quadratic step estimation. Supported parameters are:

Name	Type	Default	Description
ftolr	double in [0, inf)	0	Stop if the relative change of the criterion is below.
gtola	double in [0, inf)	0	Stop if the inf-norm of the gradient is below this value.
maxiter	uint in [1, inf)	100	Stopping criterion: the maximum number of iterations
scale	double in (1, inf)	2	Fallback fixed step size scaling
step	double in (0, inf)	0.1	Initial step size
xtola	double in [0, inf)	0	Stop if the inf-norm of x-update is below this value.

gsl

optimizer plugin based on the multimin optimizers of the GNU Scientific Library (GSL) https://www.gnu.org/software/gsl/. Supported parameters are:

Name

Type

Default

Description

eps

double in (0, inf)

0.01

gradient based optimizers: stop when |grad| < eps, simplex: stop when simplex size < eps.

iter

uint in [1, inf)

100

maximum number of iterations

opt

dict

Specific optimizer to be used.
gd:	Gradient descent.
bfgs:	Broyden-Fletcher-Goldfarb-Shann
simplex:	Simplex algorithm of Nelder and Mead
cg-pr:	Polak-Ribiere conjugate gradient algorithm
bfgs2:	Broyden-Fletcher-Goldfarb-Shann (most efficient version)
cg-fr:	Flecher-Reeves conjugate gradient algorithm

step

double in (0, inf)

0.001

initial step size

tol

double in (0, inf)

0.1

some tolerance parameter

nlopt

Minimizer algorithms using the NLOPT library, for a description of the optimizers please see 'http://ab-initio.mit.edu/wiki/index.php/NLopt_Algorithms'. Supported parameters are:

Name

Type

Default

Description

ftola

double in [0, inf)

Stopping criterion: the absolute change of the objective value is below this value

ftolr

double in [0, inf)

Stopping criterion: the relative change of the objective value is below this value

higher

double

inf

Higher boundary (equal for all parameters)

local-opt

dict

none

local minimization algorithm that may be required for the main minimization algorithm.
ln-newuoa-bound:	Derivative-free Bound-constrained Optimization by Iteratively Constructed Quadratic Approximation
ld-lbfgs:	Low-storage BFGS
ln-praxis:	Gradient-free Local Optimization via the Principal-Axis Method
ld-lbfgs-nocedal:
gn-direct-l-rand:	Dividing Rectangles (locally biased, randomized)
gn-orig-direct:	Dividing Rectangles (original implementation)
none:	don't specify algorithm
gn-direct:	Dividing Rectangles
ld-var1:	Shifted Limited-Memory Variable-Metric, Rank 1
gn-direct-noscal:	Dividing Rectangles (unscaled)
ld-mma:	Method of Moving Asymptotes
gn-crs2-lm:	Controlled Random Search with Local Mutation
ld-tnewton-restart:	Truncated Newton with steepest-descent restarting
ld-tnewton-precond-restart:	Preconditioned Truncated Newton with steepest-descent restarting
gn-isres:	Improved Stochastic Ranking Evolution Strategy
gn-direct-l-noscal:	Dividing Rectangles (unscaled, locally biased)
ln-bobyqa:	Derivative-free Bound-constrained Optimization
ld-var2:	Shifted Limited-Memory Variable-Metric, Rank 2
ld-tnewton:	Truncated Newton
ln-sbplx:	Subplex variant of Nelder-Mead
ld-tnewton-precond:	Preconditioned Truncated Newton
gn-direct-l-rand-noscale:	Dividing Rectangles (unscaled, locally biased, randomized)
ln-cobyla:	Constrained Optimization BY Linear Approximation
ln-neldermead:	Nelder-Mead simplex algorithm
gn-orig-direct-l:	Dividing Rectangles (original implementation, locally biased)
gn-direct-l:	Dividing Rectangles (locally biased)
ln-newuoa:	Derivative-free Unconstrained Optimization by Iteratively Constructed Quadratic Approximation

lower

double

-inf

Lower boundary (equal for all parameters)

maxiter

int in [1, inf)

100

Stopping criterion: the maximum number of iterations

opt

dict

ld-lbfgs

main minimization algorithm
ln-newuoa-bound:	Derivative-free Bound-constrained Optimization by Iteratively Constructed Quadratic Approximation
ld-lbfgs:	Low-storage BFGS
ln-praxis:	Gradient-free Local Optimization via the Principal-Axis Method
ld-lbfgs-nocedal:
gn-direct-l-rand:	Dividing Rectangles (locally biased, randomized)
gn-orig-direct:	Dividing Rectangles (original implementation)
gn-direct:	Dividing Rectangles
ld-var1:	Shifted Limited-Memory Variable-Metric, Rank 1
auglag-eq:	Augmented Lagrangian algorithm with equality constraints only
gn-direct-noscal:	Dividing Rectangles (unscaled)
ld-mma:	Method of Moving Asymptotes
gn-crs2-lm:	Controlled Random Search with Local Mutation
ld-tnewton-restart:	Truncated Newton with steepest-descent restarting
ld-tnewton-precond-restart:	Preconditioned Truncated Newton with steepest-descent restarting
gn-isres:	Improved Stochastic Ranking Evolution Strategy
ld-slsqp:	Sequential Least-Squares Quadratic Programming
gn-direct-l-noscal:	Dividing Rectangles (unscaled, locally biased)
ln-bobyqa:	Derivative-free Bound-constrained Optimization
ld-var2:	Shifted Limited-Memory Variable-Metric, Rank 2
ld-tnewton:	Truncated Newton
ln-sbplx:	Subplex variant of Nelder-Mead
ld-tnewton-precond:	Preconditioned Truncated Newton
gn-direct-l-rand-noscale:	Dividing Rectangles (unscaled, locally biased, randomized)
ln-cobyla:	Constrained Optimization BY Linear Approximation
auglag:	Augmented Lagrangian algorithm
ln-neldermead:	Nelder-Mead simplex algorithm
gn-orig-direct-l:	Dividing Rectangles (original implementation, locally biased)
gn-direct-l:	Dividing Rectangles (locally biased)
ln-newuoa:	Derivative-free Unconstrained Optimization by Iteratively Constructed Quadratic Approximation
g-mlsl-lds:	Multi-Level Single-Linkage (low-discrepancy-sequence, require local gradient based optimization and bounds)
g-mlsl:	Multi-Level Single-Linkage (require local optimization and bounds)

step

double in [0, inf)

Initial step size for gradient free methods

stop

double

-inf

Stopping criterion: function value falls below this value

xtola

double in [0, inf)

Stopping criterion: the absolute change of all x-values is below this value

xtolr

double in [0, inf)

Stopping criterion: the relative change of all x-values is below this value

Plugin consumers:

mia-2dgroundtruthreg mia-2dimageregistration mia-2dmultiimageregistration mia-2dmyoica-full mia-2dmyoica-nonrigid mia-2dmyoica-nonrigid2 mia-2dmyoica-nonrigid-parallel mia-2dmyoicapgt mia-2dmyomilles mia-2dmyoperiodic-nonrigid mia-2dmyopgt-nonrigid mia-2dmyoserial-nonrigid mia-3dnonrigidreg mia-3dnonrigidreg-alt mia-3dprealign-nonrigid mia-3drigidreg mia-3dserial-nonrigid

Prev	Up
Plugin type: mesh/io	Home