Re: validation from data is also insufficient

[Doug Donalson]

>
> hihi, all -

> validation from data alone is not sufficient (with enough parameters, any
> model can mimic any behavior to some degree of accuracy) - there should
also
> be some kind of model plausibility argument, based on theory or something
> (part of good modeling is making good choices of these somethings)

One of the chapters of my dissertation looks at exactly this problem.  I
took a model (published in Ecology) that was fit to a particular data set
and tested robustess of the fitted model in the presense of dempgraphic
stochasticity.  I am happy to report the the fitted model failed badly
(which was what I hoped to show.)

In one of the tests I performed, I generated 20 stochastic replicates from
an alternate model and then fitted both the Ecology authors model and the
alternate model to each stochastic replicate.  I used the author's method
for choosing the "true" model (a best fit criteria) and 7/20 times, the
Ecology authors model model was chosen as the correct model.  In comming up
with his "best fit" model, he added so much information (including an
unconstrained initial condition) that the model became a chimera, able to
look like several different systems by simply modifying the parameters.

Cheers,

   D3



                  ==================================
   Swarm-Modelling is for discussion of Simulation and Modelling techniques
   esp. using Swarm.  For list administration needs (esp. [un]subscribing),
   please send a message to <address@hidden> with "help" in the
   body of the message.
                  ==================================