fmincon pause after 2 iterations when using gradients
6 views (last 30 days)
Show older comments
Without gradients my code runs 'fine'. Meaning that it is slow to converge, but at least it is continuously iterating towards a better solution.
Adding gradients results in a freeze/pauze, usually after the second iteration. After a while it picks up again and it's better/faster than without gradients. Consecutive runs of the same script don't have this freeze/pause, which leads me to believe it's some sort of caching related issue.
'checkGradients' tells me that the gradients are good. I get my gradients from a different script by taking the symbolic jacobian of ceq, and using matlabFunction to generate a sparse optimized matlab function file.
If anyone could explain why this happens and maybe how I could resolve the slowdown I would be very grateful!
(the slowdown is not because of the 'checkGradients' as it also happens without the check. I only generate the function once, not every time I try to optimise.)
CheckGradients Information
Objective function derivatives:
Maximum relative difference between supplied
and finite-difference derivatives = 2.0994e-07.
Nonlinear equality constraint derivatives:
Maximum relative difference between supplied
and finite-difference derivatives = 4.20079e-08.
CheckGradients successfully passed.
____________________________________________________________
First-order Norm of
Iter F-count f(x) Feasibility optimality step
0 1 1.960200e+01 9.900e-01 5.822e-01
1 2 2.011323e+01 9.888e-01 1.048e+00 7.881e-01
2 3 2.174207e+01 9.884e-01 1.435e+00 4.568e-01
3 Comments
Accepted Answer
Alan Weiss
on 9 Mar 2022
I don't know for sure, but MATLAB has a just-in-time compiler that does indeed make a sort of cache for code that is repeatedly called. On first calling a function, the compiler makes a compiled version that is then reused on subsequent calls. Your matlabFunction call probably creates a long, complex function that is slow to interpret the first time, and is fast to run thereafter.
Or I could be wrong, and what is happening is your matlabFunction call creates a call to an obscure portion of the code base that takes a while to load. Sorry, I don't really know which is more likely, or how you would check it. But perhaps running your code for an iteration or two before calling your optimization will smooth things out either way.
Alan Weiss
MATLAB mathematical toolbox documentation
0 Comments
More Answers (1)
Matt J
on 8 Mar 2022
I get my gradients from a different script by taking the symbolic jacobian of ceq, and using matlabFunction to generate a sparse optimized matlab function file.
If you're repeating this process every time your constraint function is called, it's probably a bad idea. You should generate your matlabFunction once, before the optimization is launched.
See Also
Categories
Find more on Linear Least Squares in Help Center and File Exchange
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!