Download we include it here.

The PEST Control File
4-33
vector points outside of allowed parameter space (ie. above or below the upper or lower
bound at which the parameter currently resides) are sequentially frozen; the parameter
upgrade vector is then re-calculated with the frozen parameters omitted from these new
upgrade calculations. The sequence in which such parameters are frozen is important; see
Section 2.2.3 for further details. At the beginning of each new optimisation iteration, all
frozen parameters are freed. The whole process is then repeated during the next optimisation
iteration.
Sometimes, particularly in highly parameterised inversion problems, some parameters can
move to their bounds, and then back again into allowed parameter space, during the course of
the parameter estimation process. However, in most cases, if a parameter is at its bound for
more than a few optimisation iterations, it is more likely than not that the parameter is there
to stay. However PEST will continue to calculate derivatives with respect to this parameter in
order to attempt an upgrade calculation for it, even though there is a diminishing likelihood
that the parameter will ever move from its bounds. Thus one model run per iteration (two if
PEST is engaged in central derivatives calculation) is wasted. Where more than one
parameter is at its bounds the number of wasted model runs rises in proportion to the number
of such parameters.
Through use of the IBOUNDSTICK control variable, the user can now prevent this wastage
of model runs. IBOUNDSTICK is an integer variable which must be supplied as zero or
greater. If it is supplied as zero, PEST’s operation is unchanged from that described above.
However if it is set to n (where n is a positive integer), PEST will permanently “glue” a
parameter to its upper or lower bound if that parameter has been residing there since n
optimisation iterations ago. Once a parameter is “glued” to its bounds it will never move
again, for PEST will no longer include this parameter in its upgrade vector calculations. Nor
will it calculate derivatives with respect to this parameter, thus reducing the number of model
runs required per optimisation iteration.
You should carefully note the following points regarding the use of IBOUNDSTICK.
1.
If IBOUNDSTICK is set to 1, then parameters will be glued to their bounds from the
moment that they strike them (beginning at the optimisation iteration immediately
following the bounds encounter). Thus if the initial value of a parameter is at its upper
bound and IBOUNDSTICK is set to 1, then the parameter will be glued to its bound
from the very first optimisation iteration.
2.
If IBOUNDSTICK is set to 2, then 1 complete optimisation iteration will elapse (in
which the parameter is free to move back into allowed parameter space) since the
iteration at which it encountered its bound before that parameter is glued to its bound.
3.
Normally a good setting for IBOUNDSTICK is 2 to 4.
4.
Once a parameter is glued to its bound, PEST no longer calculates derivatives (ie.
sensitivities) with respect to this parameter. Where a model is nonlinear, the sensitivity
of a parameter depends not only on the current value of that parameter, but also on the
values of other parameters involved in the parameter estimation process. While PEST
has no further use of the sensitivity of a glued parameter in calculation of the parameter
upgrade vector, it does make use of that sensitivity for listing in the sensitivity file, and