fminsearch giving results below initialisation point

Hello,

I have a function that I am trying to minimise as follows

function x = kmean_per_dif_fit(img, adam)

x0 = [80,10,1];
tota = @(x)0;
totb = 0;

for i = 1:5
    err = @(x)abs(round(sum(sum((abs(img{1} - x(1))<abs(img{1} - x(2))) == 1))/x(3))-adam(i));
    tota = @(x)tota(x) + err(x);
    totb = totb + adam(i);
end

a = @(x)tota(x)/totb;

x = fminsearch(a,x0)
end

As you can see, i have three unknowns and I try to initialise them at x0 = [80,10,1]

The answers that I get yet, are not within my initilastions

  316.1776   -1.1811    2.6955

any help would be appreciated!

[John D'Errico]

Very strange. (Not what fminsearch did, but what you are doing.)

You have one function in there called tota, which is identically zero. You add it to err(x). err is something strange looking too. But it appears that err is doing something highly non-differentiable, and certainly discontinuous.

Why is that a problem? Because fminsearch is not designed to solve such a problem.

And, of course, you do not tell us what img and adam are. IMG is some arbitrary cell array, containing only cell. adam is some unknown vector, of length 5. So we cannot even test your code out to show you what it does, and why fminsearch fails.

So it is totally impossible to help you, to explain why fminseach goes where it did go.

As for why it might have done so, fminseach has NO constraints upon it. So it can freely go wherever it decides might improve the objective function.

Answer 1

[Pravin Jagtap]

Hello Udi,

I think this can happen because it depends on the function and starting point (Since 'fminseach' function uses heuristic method). The 'fminsearch' uses the Nelder-Mead simplex algorithm for getting the minimum value. Please refer to following documentation for more details on the algorithm.

https://www.mathworks.com/help/matlab/math/optimizing-nonlinear-functions.html#bsgpq6p-11