Levenberg-Marquardt not an option for LSQNONLIN

hi, I am trying to find the optimum parameters to fit a data to another.Since I have many input parameters, and when using the trust-region-reflective (default) algorithm in LSQNONLIN, the warning told me to use Levenberg-Marquardt because there must be as many equations as variables. When I am setting up the "options", my code is

%%%% options=optimoptions('lsqnonlin','Levenberg-Marquardt',.005);

and the error message is that "Levenberg-Marquardt" is not an option for LSQNONLIN. Then I tried

%%%options={'levenberg-marquardt',1e-2,'Jacobian'};

there was no error message, but there was a warning

"Warning: Struct field assignment overwrites a value with class "cell". See MATLAB R14SP2 Release Notes, Assigning Nonstructure Variables As Structures Displays Warning, for details.".

And it used trust-region-reflective algorithm again. The output of lsqnonlin didn't give me a global min, it stopped after several iterations as the step size is less than the tolerance. I don't think it used the levenberg-marquardt. What was wrong? Thank you!

Answer 1

[Alan Weiss]

What you should do depends on your MATLAB version. For R2014b or later, use the InitDamping option to pass Levenberg-Marquardt parameters. As the R2014b Release Notes describe, you set:

 options = optimoptions(@lsqnonlin,'Algorithm','levenberg-marquardt','InitDamping',0.1);

But for earlier releases, use optimset instead of optimoptions:

 options = optimset('Algorithm',{'levenberg-marquardt',0.1});

This works for R2008b and later.

Alan Weiss

MATLAB mathematical toolbox documentation

[yuese zheng]

great thank you!