Marquardt’s algorithm undertakes a derivative-based search
for the maximum-likelihood parameter estimates.
Such an algorithm or an alterative is necessary for non-linear
models.
Marquardt’s algorithm requires the values of the first and
second derivatives of the likelihood function with respect to the
current parameters.
This algorithm was implemented in SmartStats © with the ability
to automatically calculate these derivatives numerically.
However, some models have had the equivalent analytical derivatives
coded into the model.
In general, the function minimization behaviour of some, but not
all, models will be noticeably more efficient when using analytical
derivatives.
However, for complex models, calculation of analytical derivatives
is neither a practical, nor sufficiently more computationally efficient,
option to justify the work involved in calculating them without
error.
This is particularly true when you consider that most of us would
be reluctant to not verify the analytical derivatives numerically.
The reference for the implementation of Marquardt’s algorithm
in SmartStats is:
Press, W.H., S.A. Teukolsky, W.T. Vetterling and B.P. Flannery.
1993. Numerical Recipes in Fortran: The Art of Scientific Computing.
Cambridge Univeristy Press. [Numerical
recipes]
It is worth noting here that no search algorithm guarantees that
the final fit is the global minimum that is usually desired.
Sometimes the so-called ‘best-fit’ is a local minimum,
or at worst a singularity, which are useless for providing meaningful
estimates for the parameter values.
As such, practitioners may wish to choose different sets of starting
parameter values and judge whether the final estimates are consistent
among several fitting attempts.
Therefore it is always best to choose parameter values for which
you have an a priori suspicion of the final estimate.
It is with this concern in mind that, where possible, models are
parameterized using biologically meaningful and interpretable parameters
- for example, the values of polynomial coefficients usually have
no practical biological interpretation value.
Another approach to assuring the final fit is the desired fit is
to perform a systematic grid search for parameter values prior to
undertaking a simplex search.
One model offered on this website [Species
richness] is coded with this possibility.
The multidimensional grid will methodically select sets of parameter
values, usually by changing their values in proportional increments
within a reasonable (user chosen) parameter space.
The best set of parameters discovered can then be used as starting
values in a simplex search or derivative-based
search.
You can appreciate that such grid searches can be time consuming,
but they can mitigate the frustration of having to begin an estimation
with naïve parameter values.
Once SmartStats © converges on a putative ‘best-fit’,
it attempts to confirm that fit by conducting a so-called axial
search.
If this axial search (which systematically moves parameter values
away from their value at the putative minimum) discovers an improvement
in fit greater than the tolerance for convergence, the putative
fit is rejected and the search for the ‘best-fit’ continues.
|