IPOPT on steroids: fatrop solver in CasADi

Demo repo: github.com/jgillis/fatrop_demo
Fatrop is a fresh solver developen by Lander Vanroye and Wilm Decre at KU Leuven.
Exploiting the block dense structure of optimal control problems, it beats IPOPT with easily an order of magnitude.
Get up to speed with CasADi syntax, best practices and math?
* register in November Hands-On course: ocp2024.casadi.org
* help spread the word and win a Meta Quest 3 VR headset: www.linkedin.com/posts/joris-...

[@sandeepkumarr124]

Nice demonstration. I wish there was a feature within opti that can rearrange variables to obtain the best sparse structure. Compared to hpmpc, it can be expected to more robust to initial guesses which exploiting the structure as hpipm?

[@jorisgillis4017]

Note that this goes beyond hmpc/hpipm (qp solvers), this things solves nlps..
Your wish is a logical follow-up step

[@sandeepkumarr124]

What I really wanted to know was if fatrop is ipopt configured to solve optimal control problems? Are all features of ipopt retained by fatrop

[@jorisgillis4017]

​@@sandeepkumarr124 yes you could say that. Important missing features are restoration phase and limited memory hessian approx though.

[@shanshanbond7040]

This is really great. Can i try fatrop for MHE ? Also, I wish your webcam video was showing on the bottom right. Part of the code got covered.

[@jorisgillis4017]

MHE should be perfectly possible. Do check the demo repo for the code

[@andreaarlotta845]

hey, interesting work! i.y.o. could be possible to integrate a NN in the opc with fatrop? furthermore, i.y.o. using fatrop in this framework, could improve the performances wrt to casadi classic solver?
thanks a lot for the attention

[@jorisgillis4017]

You can use L4casadi together with our fatrop interface. The speedup is expected for dynamic optimization with system dynamics that are small or non-sparse (more technically, sparsity patterns that cause a large fill-in when factorized)

[@andreaarlotta845]

@jorisgillis4017 Thanks for the reply. I will try for sure your solver and your rockit toolkit! I am working with L4casadi, but with the non-real-time library (Taylor derivative), but even small MLPs take seconds or minutes for inference on Rn×1vectors.

[@alfonsosciacchitano9210]

Hey there I'm excited to use this update. Do you have this code in Matlab as well? I'm trying to follow along by coding it myself. Thanks!

[@jorisgillis4017]

I've just uploaded an .m script to the repo mentioned in the caption..

[@alfonsosciacchitano9210]

Excellent thank you!

[@RicardoTapia-l4x]

when is casadi 3.6.6 being released?

[@jorisgillis4017]

@@RicardoTapia-l4x start of august