Algorithm
files are text files that retain information on parameter definitions,
values, step sizes, etc., for specific models.
Since algorithm files are text files, you have the convenience
to edit them (carefully) using a text editor such as Notepad, though
interactive editing using 'Edit algorithm data' is typically more
efficient.
Once a particular model has been initialized, you will be required
to load an algorithm file.
You may choose to have SmartStats © create a new, default,
file (generally with all parameter values and step sizes equal to
zero, and values unfused, or you may load an already existing file.
Note that among existing algorithm files, only those having the
correct number of parameters for the initialized model will be read
by SmartStats ©, however a default file will be set to the
correct number of parameters for that model.
After clicking on 'Edit algorithm file' you can change parameter
values, their step sizes, and fuse (or link) the values of parameters.
Note that I use the term 'fuse' to avoid confusion with the typical
use of the term link in the context of a link function.
Note that parameters with a non-zero step size will be estimated,
otherwise they will be assumed to be fixed at the specified parameter
value.
The value of a non-zero step size is relevant in only two contexts.
The first context applies when using the 'simplex
algorithm' to estimate parameter values.
In this context the step size is used to calculate the initial
deviations of trial parameter values from their starting values
when undergoing this geometric search for improved parameter values.
The second context applies when numerically calculating a covariance
matrix of the estimated parameter values.
In this context the step sizes are used to scale the deviations
of the parameter values from their maximum-likelihood values during
numerical calculation of the first and second derivatives of the
likelihood with respect to the parameter values.
I rarely adjust step sizes, almost always leaving them at 10% of
the parameter value.
However, an experienced analyst may recognize situations where
tweaking the step sizes can have a beneficial effect on parameter
estimation or calculation of the covariance
matrix.
Parameter values that you wish to share a common value are fused
by co-associating them using either a one (e.g., 'A') or two (e.g.,
'BB' or 'CZ') character alphabetical code.
Fused parameter will be co-estimated and the step sizes of all
but the first in ordinal rank of the fused parameters will be fixed
at zero (0).
Parameter names (descriptions) cannot be changed by an analyst,
they are part of the compiled model code.
Other algorithm file options available are either self-explanatory
or typically unimportant for most analyses, e.g., I have never modified
the 'Step-size reduction factor'.
I have found that for maximum-likelihood estimation the 'Tolerance
for Convergence' is reliably set to ABS (absolute) with a convergence
tolerance of 0.0001.
A useful reference for further appreciating algorithm files is:
Mittertreiner, A.,
and J. Schnute. 1985. Simplex: a manual and software package
for easy nonlinear parameter estimation and interpretation in fishery
research. Canadian Technical Report of Fisheries and Aquatic Science
1384, Ottawa, Ontario, Canada.
Though most of the implementation aspects of this technical report
are out-of-date, perhaps even obsolete, the conceptual descriptions
of the simplex method, algorithm
files, and covariance calculations
are informative.
|