Extract the ADNLPModel

[dpo]

Hi guys. Congratulations on an amazing package! I'm playing around with OptimalControl.jl. If I understand well, there is an underlying ADNLPModel that can be passed to IPOPT (via NLPModelsIpopt) or to MadNLP.
I am wondering if it's possible to extract the ADNLPModel from an ocp problem. The idea would be to try different solvers.

Thanks!

[ocots]

Hi Dominique! Thanks for the message. I'm glad to read that you use the package. I still have my notes about Julia from your presentation in Toulouse few years ago :-)

Actually, you can extract the NLP model, see this tutorial.

However, we will make a new release very soon and the direct_transcription function will be replaced by another. I will tell you when it will be effective.

In the current version, you can only select ADNLPModels but in the future release there will be also the possibility to use ExaModels to model the underlying NLP model.

Note

The package OptimalControlProblems.jl can be used to get JuMP and ADNLPModel models of some optimal control problems. However, it is still in development since for some optimal control problems, both models do not correspond.

[dpo]

Hi Olivier ! Sorry for the slow reply. That works well, thank you! I especially appreciate the mechanism to recover the states, controls and costates from the ADNLPModel’s combined solution.

[ocots]

Nice!

And if one day you want to add an application of OptimalControl to the top menu, you are welcome! You simply have to follow the guidelines here. For instance, Gene regulatory networks is not in our control-toolbox organization but it is accessible from the menu.

[dpo]

Will do. Is any grid-refinement scheme implemented in OptimalControl (or another package)? If not, are there plans to add one?

[ocots]

Good point. This feature is not implemented yet but we plan to (ping @PierreMartinon, @jbcaillau)!

Here is a MRI example where I start with a grid of size 50, then I finish with 500. It helps a lot the convergence.

I don't know if it would take time to add this feature but clearly we have to!

[jbcaillau]

@dpo sure, check control-toolbox/CTDirect.jl#119 greetings from JuliaCon2025!