Next: Common optimization options, Previous: Compatibility functions, Up: Top [Index]
All frontends for optimization and for result statistics (nonlin_min, nonlin_residmin, nonlin_curvefit, residmin_stat, curvefit_stat)accept the following options, settable with (octave)optimset.
These options are handled within the frontend.
FinDiffRelStep
Column vector (or scalar, for all parameters) of fractional intervals
supposed to be used by gradient or Jacobian functions performing finite
differencing. Default: .002 * ones (size (parameters))
for
central intervals and .001 * ones (size (parameters))
for
one-sided intervals. The default function for finite differencing won’t
let the absolute interval width get smaller than
abs (FinDiffRelStep .* TypicalX
(see below).
diffp
Can be used alternatively to FinDiffRelStep
, but for central
intervals twice the specified value will be used for backwards compatibility.
diff_onesided
Logical column vector (or scalar, for all parameters) indicating the
parameters for which one-sided intervals (instead of central intervals)
should be used by gradient or Jacobian functions performing finite
differencing. Default: false (size (parameters))
.
FinDiffType
Can be used alternatively to diff_onesided
, but always applies to
all parameters at once. Possible values: "central"
(central
intervals) or "forward"
(one-sided intervals).
TypicalX
Column vector (or scalar, for all parameters) whose absolute value
specifies minimal absolute parameter values for computation of intervals
in finite differencing by gradient or Jacobian functions (see
FinDiffRelStep
). Default: 0.0001. Must not be zero.
cstep
Scalar step size for complex step derivative approximation of gradients or Jacobians. Default: 1e-20.
parallel_local
Logical or numeric scalar, default: false
. If the parallel
package, version >= 2.0.5
, is loaded, estimate gradients of
objective function and Jacobians of model function and of constraints in
parallel processes. If parallel_local
is set to an integer
> 1
, this is number of parallel processes; if it is <= 1
,
the number of processes will be the number of available processor cores.
Works for default (real) finite differences and for complex step
derivatives. Due to overhead, a speed advantage can only be expected if
objective function, model function or constraint functions are time
consuming enough. Additionally, this setting is also passed to the
individual optimization backends, which may also consider this option
(see documentation of backends). If this option is equivalent to
true
, a warning (ID: optim:parallel_local
) will be issued
if no parallel
package of a correct version is loaded.
parallel_net
Empty (default) or a parallel connections object, see function
pconnect
of the parallel
package. If not empty, estimate
gradients of objective function and Jacobians of model function and of
constraints using parallel processing in a network of machines. The
considerations regarding a speed advantage are similar to those for
option parallel_local
.
fixed
Logical column vector indicating which parameters are not optimized, but kept to their inital value.
Next: Common optimization options, Previous: Compatibility functions, Up: Top [Index]