Matlab Coder and fminsearch: what implementation strategy?

Hi all,

I'd like to use Matlab Coder to speed up optimization when the 'fminsearch' function is used. This function seems supported by Matlab R2014b (<http://www.mathworks.fr/help/coder/ug/functions-supported-for-code-generation--alphabetical-list.html)>.

Here is my problem. I do not know

1. how to apply Matlab Coder in my case since anonymous and nested functions are not supported
2. which is the best implementation strategy for the following kind of problem where I need to pass parameters and data for the optimization process

Here is the function

function [x,fval,exitflag,output] = myfun1(parameters, data_1, data_2, data_n)

    options = optimset('FunValCheck','on','MaxFunEvals',5000,'MaxIter',5000);  

    objfun = @(parameters)myfun2(parameters, data_1, data_2, data_n);

    [x,fval,exitflag,output] = fminsearch(objfun,parameters,options);

end

Thanks.

Answer 1

[Ryan Livingston]

One workaround is to model the anonymous function's parameters using a sub-function and persistent variables. The idea is to make objfun a sub-function and data_1,data_2,data_n be persistent variables in objfun. Then you use nargin of objfun to update the parameters' values when called with more than one argument.

You can see an implementation of such a function at:

http://www.mathworks.com/matlabcentral/answers/116449-use-quadgk-with-multiple-inputs-with-matlab-coder#answer_124990

that describes a similar workaround for MATLAB Coder and QUADGK.

[Nicolas]

Dear Ryan,

thank you for your prompt reply.

Your solution works and I obtain a speed improvement close to 6 for my first tries. Matlab Coder is a great tool.

Thanks again for your help,

Nicolas

[Ryan Livingston]

That's great to hear Nicolas, thanks.

You're quite welcome.