Numerical convergence error in MARK v. 9

announcements (new versions, changes, bugs, installation problems...) related to program MARK

Numerical convergence error in MARK v. 9

Postby mjh21 » Mon Nov 11, 2019 2:20 pm

Hello,

I am constructing a multi-state (recaptures only) model using the newest version (v. 9) of Program MARK. Last month, I kept getting an error saying "numerical convergence never reached" on a model structure that had previously worked. When my advisor tried running my .inp files on his computer using the exact same settings, the model ran fine. We eventually realized he was still using MARK v. 8 on his computer. I downloaded v. 8.2 and re-ran the exact same files in the same way, and they worked.

In reading through the online forum, I found this post suggesting that Windows 10 or individual computer issues may be contributing to the problem: viewtopic.php?f=2&t=3889&p=12840&hilit=version+9#p12840

However, I have tried running my model on two different Windows 10 computers, and it failed on both in MARK v. 9 and ran on both using MARK v. 8.2, regardless of if I was in "compatibility mode" or not. The same thing happened when I tried running multi-state simulations.

I wanted to report this issue to see if other people had experienced anything similar, and if there is any potential explanation for why this happened. Thanks for your help!
mjh21
 
Posts: 1
Joined: Sun Nov 10, 2019 8:25 pm

Re: Numerical convergence error in MARK v. 9

Postby cooch » Mon Nov 11, 2019 2:41 pm

mjh21 wrote:Hello,

I am constructing a multi-state (recaptures only) model using the newest version (v. 9) of Program MARK. Last month, I kept getting an error saying "numerical convergence never reached" on a model structure that had previously worked. When my advisor tried running my .inp files on his computer using the exact same settings, the model ran fine. We eventually realized he was still using MARK v. 8 on his computer. I downloaded v. 8.2 and re-ran the exact same files in the same way, and they worked.

In reading through the online forum, I found this post suggesting that Windows 10 or individual computer issues may be contributing to the problem: viewtopic.php?f=2&t=3889&p=12840&hilit=version+9#p12840

However, I have tried running my model on two different Windows 10 computers, and it failed on both in MARK v. 9 and ran on both using MARK v. 8.2, regardless of if I was in "compatibility mode" or not. The same thing happened when I tried running multi-state simulations.

I wanted to report this issue to see if other people had experienced anything similar, and if there is any potential explanation for why this happened. Thanks for your help!


There were a *lot* of changes in MARK between v. 8.xx -> v. 9.xx, including significant changes to the numerical optimzations. Without getting into the weeds too much, things that seemed to run fine in v. 8.xx might now appear to be 'having problems' in v. 9.xx, especially for parameters at/near the boundaries (in fact, you're likely to see differences in the number of parameters estimated between 8.xx and 9.xx, given what you've described). Messages about 'underflow' can usually be ignored. Warnings about 'numerical convergence not reached' are more problematic -- in point of fact, earlier versions of MARK were more forgiving (suggesting there were no convergence issues when in fact there probably were), while recent versions of MARK are more exacting -- higher numerical precision and lower convergence tolerances (meaning, it doesn't fool you into thinking convergence has been reached when it hasn't). Best thing to try is to use better starting values -- say, by starting with simple models, and using those estimates as starting values for more complicated models further up the model set. In the end, though, there may simply be no simple fix for boundary parameters.

In short -- there is likely nothing wrong with Windows 10 (at least with respect to running MARK), or compatability mode or such things. The difference is in MARK itself.
cooch
 
Posts: 1578
Joined: Thu May 15, 2003 4:11 pm
Location: Cornell University


Return to software problems/news

Who is online

Users browsing this forum: Bing [Bot] and 1 guest

cron