ADMB Mac vs PC

[Hamazaki, Hamachan (DFG)]

I am having an interesting issue that a model getting convergence in Mac but not in PC.  Same model code, data, parameter bounds, starting point.

Any ideas why this happens, and more importantly what is a solution, other than “use Mac for tis modeling task…”

 

Thanks

 

Toshihide "Hamachan" Hamazaki, 濱崎俊秀PhD

Alaska Department of Fish and Game: アラスカ州漁業狩猟局

Division of Commercial Fisheries: 商業漁業部

333 Raspberry Rd.  Anchorage, AK 99518

Phone:  (907)267-2158

Cell:  (907)440-9934

 

[Jim Ianelli - NOAA Federal]

Would need files to test if that can be reproduced

Jim

---

From: Hamazaki, Hamachan (DFG) <toshihide...@alaska.gov>
Sent: Monday, October 25, 2021 10:18:19 PM
To: us...@admb-project.org <us...@admb-project.org>
Subject: [ADMB Users] ADMB Mac vs PC

 

--
To unsubscribe from this group and stop receiving emails from it, send an email to users+un...@admb-project.org.

[Dave Fournier]

Hi,

As Jim said one needs the files to see exactly what is going on.  But this may give you a bit of perspective.  One  was a model which converged in Linux but not in Windows.  Turned out that

the default floating point round off option was different for the two OS's.

This made no substantial difference until about 200-300 function evaluations at which point the

numbers had changed enough so that the function minimizer took a different branch in the code,

and so went off in a different direction.  One found a good local minimum and the other got

stuck in a flat place.

So the first thing to do is to save all the function evaluations for your model (to 16-17 sig figs)

and see if this is what is going on.  If so it does not indicate that the Mac is superior.  It is just the luck of the draw. I was using Linux on the PC but you can compare windows the same way.

I also had a simple model which converged using Excel solver and not ADMB. Same thing I decided.

Just the luck of the minimizing path chosen. 

Of course there could be a bug in the code.  You should take the converged Mac solution

and verify that the PC version gives the same function value.

--

[James Bence]

Years ago someone sent me an excel spreadsheet where they were trying fit a model using Solver and it was not converging and asked me to look it over.   The first thing I did was refit it using Solver and it converged.  Then someone else in our group tried it on a different machine using what appeared to be the exact same version of excel (same release in those days stamped on a DVD) and the same model computer.  One converged and the other did not!  It turned out the two DVDs were released a few weeks apart and had a minor update to the solver add on that led to slightly different trajectories (even though they were labeled as the same excel release on the same date).  If we started the non-converged computer at the answer from the converged one it also identified that solution as a minimum.  Basically same thing Dave Fournier is saying.

[Johnoel Ancheta]

Is the input data in binary?

--