[Ipopt] Extract Objective Function Gradient in finalize_solution(...)

[Stefan Vigerske]

Hi,

note that there are two representations of the same NLP in Ipopt.

In your implementation of eval_grad_f you work on the TNLP that you have 
passed to Ipopt.

Internally, Ipopt works with an NLP in standardform (the one from the 
Ipopt implementation paper: no inequalities, only equations, fixed 
variables removed). The vector you get in IpoptCalculatedQuantities is 
w.r.t. this standardform. That is probably one reason why it's not so 
trivial to actually get these values.

There have been posts on this list on how to translate vectors from the 
internal NLP representation to the TNLP form. You should find these in 
the mailing lists archive, try searching for
   TNLPAdapter site:http://list.coin-or.org/pipermail/ipopt/
in google.

Stefan


On 02/01/2013 03:44 PM, Knut Heidemann wrote:
> Hi,
>
> I access the gradient values of my final objective function via
>
> /  SmartPtr<const DenseVector> curr_grad = dynamic_cast<const
> DenseVector*>(GetRawPtr(ip_cq->curr_grad_f()));//
> //  const double* grad_values = curr_grad->Values();/
>
> Unfortunately, this does not agree with the values calculated in the
> last call of "eval_grad_f".
> The first few entries do agree, but then I get values that are not
> related to the actual results.
> Furthermore, the values are not reproducable.
> For me, this looks like I am accessing values that have not been set?
>
> Might there be something wrong with the code above, or even the methods
> used therein?
>
> Kind regards,
> Knut.
>
>
>
>
> Am 18.01.2013 16:29, schrieb Stefan Vigerske:
>> Hi,
>>
>> On 01/18/2013 12:10 PM, Knut Heidemann wrote:
>>> I tried the code snippet below and get the following error:
>>>
>>> /In member function 'virtual void
>>> WLCNLP::finalize_solution(Ipopt::SolverReturn, Ipopt::Index, const
>>> Number*, const Number*, const Number*, Ipopt::Index, const Number*,
>>> const Number*, Ipopt::Number, const Ipopt::IpoptData*,
>>> Ipopt::IpoptCalculatedQuantities*)'://
>>> //error: invalid use of incomplete type 'struct
>>> Ipopt::IpoptCalculatedQuantities'//
>>> //error: forward declaration of 'struct
>>> Ipopt::IpoptCalculatedQuantities'/
>>>
>>> Can this be due to further missing #includes or what kind of problem do
>>> I face here?
>>
>> Yes.
>> You need to include the header file where the
>> IpoptCalculatedQuantities class is defined.
>>
>> Stefan
>>
>>>
>>> Best,
>>> Knut.
>>>
>>> Am 17.01.2013 19:21, schrieb Hans Pirnay:
>>>> this should work:
>>>>
>>>> SmartPtr<DenseVector> curr_grad = dynamic_cast<const
>>>> DenseVector*>(GetRawPtr(ip_cq->curr_grad_f()));
>>>> const double* grad_values = curr_grad->Values();
>>>>
>>>> You'll probably have to #include "IpDenseVector.hpp" and maybe some
>>>> other dependencies.
>>>>
>>>> Hans
>>>>
>>>> On Thu, Jan 17, 2013 at 6:30 PM, Knut Heidemann
>>>> <heidemannknut at gmail.com> wrote:
>>>>> Hi,
>>>>>
>>>>> I am trying to access the objective function gradient information
>>>>> after my
>>>>> problem has been solved.
>>>>> I assume that this is done in finalize_solution(...).
>>>>> Unfortunately, I do not manage to use "ip_cq" for this purpose. Can
>>>>> anybody
>>>>> tell me how the appropriate line of code would look like?
>>>>>
>>>>> Best regards,
>>>>> Knut Heidemann.
>>>>> _______________________________________________
>>>>> 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
>>>
>>
>
>