[Ipopt] Runtime modification of the model

[Paul Smith]

On Sun, Feb 27, 2011 at 4:57 PM, Horand Gassmann <Horand.Gassmann at dal.ca> wrote:
>> Thanks, Horand, for your reply. I understand that Ipopt can solve, in
>> the same C++ program, a bunch of optimization problems if one knows
>> the full details of the model before we run it -- one can feed the
>> program through files or through user interaction, as you correctly
>> suggest. However, in my case, the details of the model cannot be known
>> before the C++ program runs, as the parameters and the number of
>> constraints and variables are determined while the program is running
>> (the number of variables, the number of constraints, etc., depend on
>> random number generated, during execution, by the C++ program).
>> Moreover, the structure of the current model (in the current iteration
>> of the loop) also depends on the solution of the model solved in the
>> previous loop iteration. Thus, what I am needing is to implement is
>> the following program flow:
>>
>> LOOP BEGIN
>>  1. Clear the model;
>>  2. Generate some random numbers;
>>  3. Calculate the structure of the new model according to the random
>> numbers generated in 2 and according to the solution obtained for the
>> model solved in the previous loop iteration;
>>  4. Setup the model according to the structure calculated in the previous
>> step;
>>  5. Solve the model;
>>  6. Save the solution of model;
>> LOOP END
>
> Right. So what's the problem? Doing it this way, it makes no difference
> whether the problem you solve is deterministic or stochastic. You get the
> problem dimensions, then you populate the TNLP object, and off you go.
> Presumably you know how to do that with fixed problem dimensions; the
> stochastic part should be straighforward. You generate the model and data
> /before/ you set up the instance.

Thanks, Horand. The issue is how to populate the TNLP object.
According to the Ipopt examples, we need to write the following
methods:

    * Method get_nlp_info
    * Method get_bounds_info
    * Method get_starting_point
    * Method eval_f
    * Method eval_grad_f
    * Method eval_g
    * Method eval_jac_g
    * Method eval_h
    * Method finalize_solution

It seems to me that one needs to know in advance the structure of the
model to write theses methods. For instance, suppose that one wants to
write the method eval_f for the following objective function:

x^2 + b y^3,

with b being extracted from an uniform distribution. How could one
write the eval_f method?

But perhaps there a different way of populating the TNLP object that I
am now aware of.

Paul