I just would like a little bit more of insight in how automatic calibration without Kalman filtering is done in Vensim. I assume some type of nonlinear optimal search algorithm is used, taking as starting values the nominal ones provided. I would like to know, for example, why the order of the parameters to be calibrated is important? And given that this is so, how should the order be determined? Also, given that I have, let's say only 10 points of data where to make the comparison, what is the upper limit of parameters that reasonably can be calibrated?
Thanks in advance for your answers,
automatic calibration on Vensim
-
- Senior Member
- Posts: 1107
- Joined: Wed Mar 12, 2003 2:46 pm
The optimization used is a variation of the Powell Search method - see the bibliography in the reference manual. Order sometimes matters, but not in any discernible pattern. Using RRandom restarts should take care of that.
If you have 10 data points you might be able to calibrate 2 parameters in a meaningful way. Under special conditions you might be able to do more but it would be hard to justify unless some definitional relationships must hold (in which case automated calibration is probably not appropriate).
If you have 10 data points you might be able to calibrate 2 parameters in a meaningful way. Under special conditions you might be able to do more but it would be hard to justify unless some definitional relationships must hold (in which case automated calibration is probably not appropriate).