Calling CppAD::ipopt::solve from CppADCodeGen about cppadcodegen HOT 8 OPEN

Hi,

Is there any more info I can provide to help with this question?
Thanks,
Andy

from cppadcodegen.

Hello Andy,

From what I can see from CppAD's support for Ipopt this would require a re-implementation for CppADCodeGen since CppAD uses AD::ADFun and CppADCodeGen uses different types and requires additional steps (compilation step).
Unfortunately, there is no simple way to just start using CppAD's support for Ipopt.

Kind regards,
João

from cppadcodegen.

I think that working on a faster than CppAD::ipopt::solve interface to ipopt, along the lines suggested above, is worth while but not a simple effort. If someone is interested in taking the lead on this, I would be willing to will to help.

from cppadcodegen.

Thank you for the advice.

from cppadcodegen.

@bradbell and @joaoleal what would be the first and simplest step to make progress on using cppadcodegen for ipopt via cppad?

How do we make progress towards converting the example I first linked to so it can work with cppadcodegen?

from cppadcodegen.

I suggest you start by programming up the simple ipopt example
https://github.com/bradbell/cppad_mixed/blob/master/example/ipopt_run.cpp
And modifying it so that the derivative functions are calculated using CppADCodeGen.
Once you have done that, you can extend it to work with an arbitrary function.

You will note that the example above is documented using omhelp.
I am working on a program that converts omhelp to sphinx (I do not plan to continue supporting omhelp).
I expect the omhelp to eventually be converted to the following extension of sphinx:
https://bradbell.github.io/xrst/doc/user_guide.html#purpose

from cppadcodegen.

@bradbell thanks for the suggestion. @joaoleal can you suggest a good cppadcodegen example that will help with this task?

from cppadcodegen.

There is an example that creates a model for a Lagrangian here:
https://github.com/joaoleal/CppADCodeGen/blob/master/example/lagrangian.cpp

I would, however, not do it like this. Nevertheless, it can be used as a starting point.

I would instead create two models:

• a model for the objective function
• a model for the NLP with the objective function and the constraints (to determine Lagrangian-related information)

For the NLP model I would configure a custom sparse hessian sparsity with only half of the Hessian (to improve performance).
You can also get away with only the NLP model but there are cases where there is no need to compute information related to the constraints (e.g., eval_grad_f).

If there are any parameters in the model I would also remove them from the sparsity patterns.

from cppadcodegen.