[Developers] your post about admb convergence

Ian Taylor ian.taylor at noaa.gov
Mon Oct 22 13:07:55 PDT 2012


I added the idea of a text output of gradients as Issue
#109<http://www.admb-project.org/redmine/issues/109>,
but don't have time or skill to figure out how to do this myself.

On Sun, Oct 21, 2012 at 9:40 PM, Mark Maunder <mmaunder at iattc.org> wrote:

>  See below****
>
> ** **
>
> *From:* Liz Brooks [mailto:liz.brooks at noaa.gov]
> *Sent:* Thursday, October 11, 2012 11:19 AM
> *To:* Mark Maunder
> *Subject:* Re: your post about admb convergence****
>
> ** **
>
> hi mark,
>
> thanks for the reply to my question.  it was motivated by a more generic
> question about whether one can simply look at the gradient and assert that
> a model has or hasn't converged.  even if the magnitude is not 1e+18 (as in
> the example online), but say rather that 98% of the parameters have a final
> gradient between 0.1 and 10, i would argue that one can conclude the model
> has not converged, particularly when the additional screen info indicates
> the model cannot improve and is exceeding the number of iterations, and the
> user-defined convergence criteria is 1e-5.  you may or may not want to
> weigh in on this more general question, but feel free to.
>
> related to this point, i have a pretty brute force way of saving the final
> matrix of gradients by saving a screen dump in R and then parsing all of
> that saved text into a matrix.  who would i have to make a request to at
> 'ADMB Headquarters' to have the final gradient information printed to a
> text file?
>
> cheers, and thanks again,
> liz
>
> ****
>
> On Thu, Oct 11, 2012 at 1:35 PM, Mark Maunder <mmaunder at iattc.org> wrote:*
> ***
>
> Liz,****
>
>  ****
>
> That model has definitely not converged due to the high gradient. This
> type of very high gradient is usually caused by using the  += sign to
> assign something to the objective function without setting the objective
> function (or some other parameter, i.e. something used in the posfun
> function)  to zero at the start of the procedure section.****
>
>  ****
>
> Regards,****
>
>  ****
>
> Mark****
>
>  ****
>
> *From:* Liz Brooks [mailto:liz.brooks at noaa.gov]
> *Sent:* Thursday, October 11, 2012 6:58 AM
> *To:* Mark Maunder
> *Subject:* your post about admb convergence****
>
>  ****
>
> hi mark,
>
> i was reading through some of the replies to questions on the ADMB website
> pages, and saw this one that both you and dave had replied to.  not sure if
> you remember it.  i noticed that the convergence criterion specified was 1
> e-12, yet all of the gradients are enormous (5.7e+18 is the largest, 1e+6
> is the smallest).  the minimization obviously stopped because it exceeded
> the number of function evaluations and wasn't making progress:
> ic > imax  in fminim is answer attained ?
>
> however, if one notes the magnitudes of the gradients, wouldn't you
> conclude that the model had not attained a minimum?  i've pasted the screed
> dump from the original question below (and the weblink).  i'm curious how
> the gradients can be so large yet, as i understand the thread, the
> parameter estimates were at the "true" solution values.  seems
> paradoxical.  do you have any insights?
>
> thanks
> liz
>
>
> https://groups.google.com/forum/?fromgroups=#!topic/admb-users/74AA16IWQNk
> Sylvain Bonhommeau
>
> 25 variables; iteration 740; function evaluation 907
> Function value  -2.7878808e+05; maximum gradient component mag   5.7352e+18
> Var   Value    Gradient   |Var   Value    Gradient   |Var   Value
> Gradient
>   1-36.23879  4.58537e+02 |  2  0.00190 -3.50573e+18 |  3  1.30460
> 2.70468e+16
>   4  1.37629  7.65838e+15 |  5  0.01472  5.73518e+18 |  6  0.00288
> 1.15859e+18
>   7 -0.03205  1.47477e+17 |  8 -0.52865 -4.26941e+14 |  9 -0.52865
> 1.04332e+15
>  10 -1.47135 -6.57274e+14 | 11  0.12112 -1.61428e+16 | 12  0.16213
> -2.80633e+16
>  13  0.08573 -1.75030e+16 | 14 -0.00369 -1.96681e+16 | 15 -0.21428
> 7.12243e+15
>  16-11.81483  1.82708e+14 | 17  4.18517  2.72772e+14 | 18  1.81483
> 3.84233e+13
>  19  0.21166  3.25505e+13 | 20  0.09244  1.03861e+16 | 21  0.05250
> 1.08502e+15
>  22  0.26054  2.77481e+15 | 23  0.78956 -9.50897e+13 | 24 210.1177
> 2.82176e+09
>  25 227.8695 -9.76063e+06 |
>   ic > imax  in fminim is answer attained ?
> Var   Value    Gradient   |Var   Value    Gradient   |Var   Value
> Gradient
>   1-36.23882  4.58537e+02 |  2  0.00190 -3.50573e+18 |  3  1.30460
> 2.70468e+16
>   4  1.37629  7.65838e+15 |  5  0.01472  5.73518e+18 |  6  0.00288
> 1.15859e+18
>   7 -0.03205  1.47477e+17 |  8 -0.52865 -4.26941e+14 |  9 -0.52865
> 1.04332e+15
>  10 -1.47135 -6.57274e+14 | 11  0.12112 -1.61428e+16 | 12  0.16213
> -2.80633e+16
>  13  0.08573 -1.75030e+16 | 14 -0.00369 -1.96681e+16 | 15 -0.21428
> 7.12243e+15
>  16-11.81483  1.82708e+14 | 17  4.18517  2.72772e+14 | 18  1.81483
> 3.84233e+13
>  19  0.21166  3.25505e+13 | 20  0.09244  1.03861e+16 | 21  0.05250
> 1.08502e+15
>  22  0.26054  2.77481e+15 | 23  0.78956 -9.50897e+13 | 24 210.1177
> 2.82176e+09
>  25 227.8695 -9.76063e+06 |
> Function minimizer not making progress ... is minimum attained?
> Minimprove criterion =   0.0000e+00
>
>  - final statistics:
> 25 variables; iteration 741; function evaluation 937
> Function value  -2.7879e+05; maximum gradient component mag   5.7352e+18
> Exit code = 1;  converg criter   1.0000e-12
> Var   Value    Gradient   |Var   Value    Gradient   |Var   Value
> Gradient
>   1-36.23882  4.58537e+02 |  2  0.00190 -3.50573e+18 |  3  1.30460
> 2.70468e+16
>   4  1.37629  7.65838e+15 |  5  0.01472  5.73518e+18 |  6  0.00288
> 1.15859e+18
>   7 -0.03205  1.47477e+17 |  8 -0.52865 -4.26941e+14 |  9 -0.52865
> 1.04332e+15
>  10 -1.47135 -6.57274e+14 | 11  0.12112 -1.61428e+16 | 12  0.16213
> -2.80633e+16
>  13  0.08573 -1.75030e+16 | 14 -0.00369 -1.96681e+16 | 15 -0.21428
> 7.12243e+15
>  16-11.81483  1.82708e+14 | 17  4.18517  2.72772e+14 | 18  1.81483
> 3.84233e+13
>  19  0.21166  3.25505e+13 | 20  0.09244  1.03861e+16 | 21  0.05250
> 1.08502e+15
>  22  0.26054  2.77481e+15 | 23  0.78956 -9.50897e+13 | 24 210.1177
> 2.82176e+09
>  25 227.8695 -9.76063e+06 |
>
> --
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>
> Liz Brooks, PhD
> Operations Research Analyst
> NOAA/NMFS
> Northeast Fisheries Science Center
> 166 Water Street                       phone: 508.495.2238
> Woods Hole, MA  02543             fax: 508.495.2393
>
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~****
>
>
>
>
> --
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>
> Liz Brooks, PhD
> Operations Research Analyst
> NOAA/NMFS
> Northeast Fisheries Science Center
> 166 Water Street                       phone: 508.495.2238
> Woods Hole, MA  02543             fax: 508.495.2393
>
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~****
>
> _______________________________________________
> Developers mailing list
> Developers at admb-project.org
> http://lists.admb-project.org/mailman/listinfo/developers
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.admb-project.org/pipermail/developers/attachments/20121022/934d950d/attachment.html>


More information about the Developers mailing list