Yet most models, in thought or in code, need to be simplified versions of the systems they describe. On some level. Otherwise, you end up in the impractical situation with a map as big as the world it describes, and no overview from which to draw conclusions.
The vast majority of useful models therefore lie somewhere on the scale between those two extremes. Neither perfectly round abstract balls, nor perfect liknesses of a cow in all its irrgular glory, but a compromise - they're slightly aspherical cows.
*****
I've been thinking about modelling since 2003-2004, when I did my MSc project and studied a handfull of models supposed to describe the same system. I compared them to each other and to experimental data. It was... illustrative, and rather depressing, how little any of them resembled each other. And how hard it was to find decent data to back up any parameter values, not to mention how very hard it was to try to reimplement that model in a parallel version based on C and MPI without introducing any bugs.
I then progressed to spending a very frustrating year trying to take a model from one simulator and implement it in another - they were supposed to behave the same, but due to hardcoded simulator differences and a few obscure bugs there was no way to reach that goal completely. Finally getting close, I realised there was no good model for the input my modeled biological system was supposed to be evolutionarily adapted to process - beyond the commonly used 'let's hit the system with a lot of Gaussian noise'. And onward...
