|
Comments to Preliminary Draft Report (First Version)
Contribution by Victor M. Ponce, Panelist
July 8, 2005
Response to comments by Norm Jones on July 8, 2005.
- Jones is not sure how helpful the following statement is:
"There are three types of errors in modeling: (1) numerical errors, which
are caused by roundoff and/or truncation, (2) physical errors, attributed to
inaccurate parameter estimation, and (3) errors that are traceable to poor
data quality. Calibration and validation examples using RSM should identify
these three sources of errors. Numerical errors can be minimized by a
judicious choice of grid resolution, physical errors can be minimized by the
proper choice of parameter ranges, and data-quality errors can usually only
be assessed in a qualitative way, however, their importance cannot be
overemphasized. Full-model validation demands the explicit separation of
errors; otherwise, one could be calibrating numerical errors against
physical and/or data-quality errors. The validation procedure should take
into account the following considerations: (1) To the extent possible,
eliminate the numerical errors; (2) calibrate to the expected values of the
physical parameters; and (3) If necessary, assess the quality of the
measured data."
Specifically, Jones states: "To say that [the statement] 'physical errors can be minimized by the
proper choice of parameter ranges' implies that finding an optimal set of parameter
values is as simple as ensuring that proper limits are set on the parameter values."
I see no great problem with the statement. It says that "physical errors can be minimized ..."
It does not say "physical errors can be eliminated ..."
If deemed necessary, a compromise statement might be: "physical errors can be reduced ..."
As Jones points out, the lack of parameter observations (that is, measurements)
limits the parameter estimation, but the fact remains that applicable parameter ranges need to be established a priori.
Disregarding this practice can result in mathematical estimations outside of the appropriate range,
which may be good calisthenics, but is invariably bad modeling. There are a number of examples in practice
where parameters have been estimated outside of their physically realistic ranges while pursuing sophisticated
"parameter estimation techniques." One that comes to mind is that of the
U.S. Army HEC's HEC-1 (Hydrologic Engineering
Center, 1979, 1985, 1990), where the kinematic wave scheme's
uncontrolled numerical diffusion was matched to the observed physical diffusion.
This was accomplished through the estimation of parameters such as Manning's n often outside of its applicable
range.
|