|
by Victor M. Ponce July 13, 2005
I agree with you that it would be unwise to "hard-wire" the expected parameter ranges into the modeling code. Barring this, there appear to be two options: (1) the John Doherty approach, where the user is allowed to enter the minimum and maximum values for each parameter, and (2) the Mary Hill approach, which would leave the user to decide for him/herself the appropriateness of estimated individual parameter values. It seems to me that the first approach is automated, or automatical. In this approach, the model is helping the user to determine whether the parameter estimation remains within realistic bounds. If an actual application reveals that some, or many, of the parameters are indeed at the imposed limits, the user can make a local assessment and take appropriate measures. The second approach is manual. Since no limits are set a priori, the parameters can conceivably take any value, depending on the problem and its relation to the "automated parameter estimation utility" being used. The burden of proof shifts to the user, who must manually assess any inconsistencies and out-of-range values in the estimated parameters. The choice between approaches remains one of preference. For routine applications, I prefer the first approach, because it is safer. However, I can see how some users may want to see how the model behaves freely, without any artificial restrictions imposed on the parameter ranges. A workable compromise may be a model where the user is allowed, at his/her choice, either to impose or not impose limits on the parameters. Then, both camps would be happy. Norm, thanks for being proactive in discussing these issues, which help clarify modeling strategy. |
| 050713 |