[Ipopt] Good solutions with errors in derivative test

Stefan Vigerske stefan at math.hu-berlin.de
Wed Feb 1 22:45:33 EST 2017


there is a jacobian_approximation option in the TNLPAdapter. I think 
it's not part of the options documentation because it is strongly 
advised to provide exact first derivatives, often computed by some AD 
code. For Hessians, there is hessian_approximation,


On 02/01/2017 09:54 PM, Chunhua Men wrote:
> Since IPOPT has derivative test, is it possible to use it to compute
> derivatives (in grad_f, jac_g and hess), instead of implementing them by
> our own?
> Chunhua
> On Wed, Feb 1, 2017 at 6:22 AM, Andrew Spiteri <andrew.spiteri at um.edu.mt>
> wrote:
>> Hi all,
>> I am using IPOPT (MATLAB interface) to solve trajectory optimisation
>> problems of medium complexity (roughly 100 variables with 300 constraints).
>> Due to this complexity, I am using INTLAB to automatically calculate the
>> gradient, Jacobians and Hessians, however I am now noticing that the IPOPT
>> derivative test is showing a considerable number of errors (in grad_f,
>> jac_g and hess), with relative errors ranging from about 0.0001 to 1. I'm
>> not sure why INTLAB seems to be giving inaccurate values, but I'm
>> suspecting the various vector manipulations (scaling, splitting and
>> merging) which occur in eval_f, eval_g and eval_hess.
>> The confusing part is that the solutions actually appear as expected,
>> despite these errors being present. Furthermore, solutions are typically
>> found within 100 iterations, which seems to indicate a reasonably good
>> performance. So my question is: is it possible to have good solutions with
>> good performance when the aforementioned errors are present in the problem?
>> Thanks in advance.
>> Andrew Spiteri
>> _______________________________________________
>> Ipopt mailing list
>> Ipopt at list.coin-or.org
>> http://list.coin-or.org/mailman/listinfo/ipopt
> _______________________________________________
> Ipopt mailing list
> Ipopt at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/ipopt


More information about the Ipopt mailing list