Optimisation Crash

[rahuliim]

I am running Vensim Pro 6.0b under Windows XP. I am trying to calibrate a simple two stock (Potential Adopter & Adopter) diffusion model. But every time I try optimising it, the system crashes. Any insights?

[tomfid]

We need to know a little more - can you post the model and your control files? Upgrading to 6.1 could help, though I don't recall any optimization specific bug fixes.

[rahuliim]

Hello Tom,
The model files are attached in a ZIP file. I upgraded to 6.1c. The incidents are fewer but still it crashes.
Thanks for the help.

Rahul

[tomfid]

This works fine for me on Win7 6.1c - I assume you're on a Mac though? I'll move it over and give that a try.

One thing that would help quite a bit here is to put upper and lower bounds on all your variables. Even if you can't determine narrow limits a priori, searching from 0 to 1 billion is a lot better than searching from 0 to infinity.

Tom

[tomfid]

Actually I see that you're using XP so I tried that and I can replicate the crash. It works fine on Win7 and Mac.

Your model generates floating point errors in its base run. Resolving those (try a much smaller time step for one thing), plus adding parameter bounds, will solve the problem, though it shouldn't be happening.

[rahuliim]

Thanks Tom.
I did suspect XP as the culprit. My student ran it on Mac and there was no problem. I shall try fix the bounds. Hopefully that would settle it. But is you optimizer assured to give global optima? I would like to know more. Could you direct me to a good resource?

Rahul

[tomfid]

The optimizer is only statistically likely to find the optimum if you use multiple start for long enough. This is true of all global optimization approaches that can't exploit problem-specific information. If your payoff surface is rough (many local optima), other approaches might be better, including the SA variant of the new MCMC algorithm ( http://www.vensim.com/documentation/mcmc_sa.htm ). However, this is unlikely to be necessary for a simple calibration problem.