Fminsearch high dimensional optimization (Excessive "Reflect" in procedure)

Hi,

I am currently trying to optimize a fairly high dimensional nonlinear objective function (n=25 to 35) using the nelder meade method-fminsearch.

As I am running the program there are excessive "reflect" procedures (very rare "contract") in each iteration which is causing the iteration to max out without convergence.

Would anyone have any tips with regards to this problem? I've tried loosening the tolerances (TolX=1e-2 TolFun=1e-3) with no luck.

Lastly, using a gradient based algorithm would be not feasible in my case because i have a need to numerically compute the gradient that is quite large.

Any help would be appreciated.

Thanks!

Answer 1

[John D'Errico]

Tip? Using fminsearch on a function with 25-35 variables is a fool's task. Sorry, but it is. Don't shoot the messenger. That it is having problems is a reflection of that fact.

Answer 2

[Sargondjani]

im facing a similar problem, and im looking into the global optimization toolbox. patternsearch or the genetic algorithm might be helpful