-
|
I am using Ipopt to solve a model predictive control problem with the C++ interface. I save the output in the callback function ma27: ma57: ma97: |
Beta Was this translation helpful? Give feedback.
Answered by
svigerske
Feb 17, 2022
Replies: 1 comment 3 replies
-
|
I guess there are some bad values (nan, inf, very tiny, very large) in some matrix or vector so that the linear system cannot be solved. You might want to try setting the print_level really high to see where these values come from. Having a smaller problem to reproduce the problem could be handy. But first, you may want to enable option check_derivatives_for_naninf to check that your evaluation routines do not accidentally pass nan's or inf's to Ipopt. |
Beta Was this translation helpful? Give feedback.
3 replies
-
|
Thank you for your suggestion. I checked the max and min values of the parameters returned by the interface function and they all seem reasonable. Since NLP is exported from autodiff, there are some tiny (1e-10 scale) non-zero entries in hessian and jacobian. I've tried rounding them to a certain precision to avoid such small values, but it didn't help with the problem. I have Finally, I tried increasing the My understanding (perhaps wrong) is that maybe the KKT systems of my NLP are ill-conditioned under certain conditions. According to Short Tutorial: Getting Started With Ipopt in 90 Minutes, Ipopt tries to perturb the hessian to increase the success rate. Any suggestions on how to fix this? And what options for hessian perturbation might improve the condition of the KKT system? Thanks! |
Beta Was this translation helpful? Give feedback.
-
|
Yes, if there are problems with solving the linear system ((6) in the tutorial), then Ipopt increases the delta_x term to force the matrix to become positive definite. You see this in ma27.log and ma57.log where delta_x gets increased until 10^22. At that point, Ipopt gives up. As you see from (6) and (7) in the tutorial, the starting point influences W_k and determines X_k and Z_k directly. Since you reduce the values for the various |
Beta Was this translation helpful? Give feedback.
-
|
I turned up the print_level and found that the warm start point from the interface had bad values. Thank you for your help! |
Beta Was this translation helpful? Give feedback.
Answer selected by
yanhaoy
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
I guess there are some bad values (nan, inf, very tiny, very large) in some matrix or vector so that the linear system cannot be solved. You might want to try setting the print_level really high to see where these values come from. Having a smaller problem to reproduce the problem could be handy.
But first, you may want to enable option check_derivatives_for_naninf to check that your evaluation routines do not accidentally pass nan's or inf's to Ipopt.