topic Re: OPTMODEL : Blackbox Solver in Mathematical Optimization, Discrete-Event Simulation, and OR

OPTMODEL : Blackbox Solver

ODENWALD — Mon, 07 Dec 2020 09:20:13 GMT

Hi:

For the first time in my life I used the new Blackbox Solver on a moderately non-linear problem.

I started between 10 and 20 GAs.

In any attempt the result was Failed with numbers for Infeasibility which seemed rapidly going in a kind of stalling.

Also I could not detect a result of the say 10 global GAs whence I had expected to somehow "see" different partial results from differing global GAs as e.g. a kind of jumping in the values of Infeasibilty.

I got not one single feasible solution.

It appeared that all the GAs searched in a false region of the search space.

Is there a way to throw out the GAs into a wider range of the search space ?

If you have some vague idea where a good solution could be how can I use that info as kind of starting values for some variables (which I teach since ages are very important in solving non-linear problems) ?

Is there a hidden option to the end to tell the solver please find me just a feasible solution ?

I was a bit disappointed when comparing GAMS with a MINLP-solver with Blackbox.

Kind regards,

Odenwald.

Re: OPTMODEL : Blackbox Solver

StevenGardner — Mon, 07 Dec 2020 21:12:58 GMT

There are some strategies that might help find feasible solutions in this case. It would be much easier to know which strategies might help if I could see your code and understand exactly what type of problem you're trying to solve. Is there any way you could share your code?

In any case, the black-box solver should always attempt to find feasible solutions before trying to improve the objective value. It seems as if it was never able to find any feasible solutions in your case. But again, it's hard to really know what's going on without more specific information about your setup.

Re: OPTMODEL : Blackbox Solver

ODENWALD — Tue, 08 Dec 2020 11:15:04 GMT

Hi Steven :

Unfortunately I cannot present the code publicly.

We are working on linearization , seems to work ; we get solutions that are worth being discussed, albeit very wide bounds.

We hope to get forward.

I ran the more or less toy examples that come with Blackbox. That worked. I tried 3 real ones and in any case obtained "Failed" in all variants that I tried. I'm far from saying that my cases are the world but SAS publishing more advice on how to effectively use Blackbox could be helpful.

Rob Pratt recently remarked that some kind of automatic linearization would be coming soon (soon is my formulation).

In the 1990s SAS has written quite some procedures to attack or solve non-linear statistical problems that used an exact or approximate linearization to get a. feasible points and/or b. good starting points for non-linear iterations.

It would be great if that strategy could be implemented for Blackbox, too.

Odenwald .