Page 1 of 1

Replicate optimization

Posted: Wed May 22, 2019 7:26 am
by portak
Hello,

Is there a reason why I cannot replicate the optimization results that I find with the .out file in the attachments?

Thank you.

Re: Replicate optimization

Posted: Wed May 22, 2019 7:55 am
by Administrator
You'll need to provide a lot more information in order for us to help.

Re: Replicate optimization

Posted: Wed May 22, 2019 8:21 am
by portak
I use the same payoff function and same optimization parameters with the same min and max values. So I think that using the same Optimization control parameters, I should have exactly the same results. However, this is not the case.

1) Does the optimization output files contain the payoff function that I use for it?
2) I use a vector multiple start. If I understood correctly, each parameter is evaluated (one at a time) between its min and max value. Is it possible that the difference is because of this uniform search between its min and max? I do not fix a seed and maybe the tolerance level that I set is really tolerant.

Is there a reason to not expect the same outputs, while using the same payoff, same parameters and same controls?

Re: Replicate optimization

Posted: Wed May 22, 2019 8:49 am
by Administrator
If you are using the same control files on the same model with exactly the same starting data, the results should be identical. If not, upload the model and all supporting files with full instructions on how to create the two different sets of results and we can take a look.

Re: Replicate optimization

Posted: Wed May 22, 2019 9:03 am
by portak
Maybe it's because of the starting data. I'll check that. Thank you!

Re: Replicate optimization

Posted: Wed May 22, 2019 11:25 am
by portak
I figured out that it was because of order of the parameters. The parameters were the same but the order was different.

Re: Replicate optimization

Posted: Wed May 22, 2019 1:51 pm
by tomfid
Order matters, though in a sense that's a bad sign - it means there are many optima and you haven't found the global max. Using the RRandom multiple start helps somewhat because it varies the search order.

The random seed for the optimizer is another factor.

Re: Replicate optimization

Posted: Fri May 24, 2019 7:21 am
by portak
Yes, indeed I have many optima. I now try with rrandom multiple start but I guess I will never be sure of finding the global optimum.
I am using Powell optimizer. Do you think it is the right choice for my case (multiple optima)?

Re: Replicate optimization

Posted: Sun May 26, 2019 3:55 pm
by tomfid
Yes. You can also use MCMC for simulated annealing, but I always start with random multistart (use the RRandom variant).

Re: Replicate optimization

Posted: Thu May 30, 2019 7:05 pm
by portak
Is it possible for you to suggest me some exemplary papers that uses Vensim optimization ?

Re: Replicate optimization

Posted: Thu May 30, 2019 7:34 pm
by tomfid
Are you looking for applications, or how-to-use?

Re: Replicate optimization

Posted: Thu May 30, 2019 7:41 pm
by portak
Mostly applications.

Re: Replicate optimization

Posted: Fri May 31, 2019 10:03 pm
by tomfid
This is a good one from the 2015 SD conference:

Jalali, Mohammad with Armin Ashouri Rad, Oscar Herrera-Restrepo and Hui Zhang Information Diffusion through Social Networks: the Case of an Online Petition (Dana Meadows Award Winner)

The supporting material includes the full model and a command script to run it, as I recall.

Re: Replicate optimization

Posted: Fri May 31, 2019 10:06 pm
by tomfid
A couple more that should be replicable:

https://metasd.com/category/model-libra ... imization/

Re: Replicate optimization

Posted: Sat Jun 01, 2019 4:13 am
by portak
Thank you!