fminsearch with part of input function

hello,

i have defined a function that is model, and want to use a fminsearch for a first value for the lsqnonlin,but i get all the time "Assignment has more non-singleton rhs dimensions than non-singleton subscripts" also in the error it says fv(:,1) = funfcn(x,varargin{:}); the model_error is the output to optimise

my code:

        param0=[1, 50];
        param0=fminsearch(@(param) Model(param,t,inlet_signal,outlet_signal),param0);
function [error_model,h]= dispmodel_2param(param,t,x,y)
            theta = t/param(2);
            h = (sqrt(param(1)./(4*pi*theta)).*exp(-param(1)./(theta*4).*(1-theta).^2));
            h(1) = 0;
            y=[y;y(end)*zeros((length(E_theta)+length(x)-1)-length(y),1)];
            error_model= y-conv(h/param(2),x);
end

Answer 1

[Walter Roberson]

No matter whether t, inlet_signal and outlet_signal are scalars or vectors, your line

            y=[y;y(end)*zeros((length(E_theta)+length(x)-1)-length(y),1)];

is making y into a vector at that point. Then your error_model would be a vector because it involves y.

When you use fminsearch there is a single output for the function, and it must return a scalar.

The form you are using appears to have been designed for use with lsqnonlin, which expects the individual predictions to be returned rather than a sum-of-squares . lsqnonlin internally does the equivalent of sum( (fun(Param) - y).^2 ) but fminsearch does not.

[franco otaola]

thanks for the answer,

inlet_signal, outlet_signal, are vectors.

yes, you are completly correct that it was designed for lsqnonlin, as it is how i was doing my solve, but the issue i had was that the lsqnonlin was somewhat dependant of the initial values, so i wanted to use fminsearch to find estimations to be able to have "good" param0 values for the lsqnonlin.

from what you are telling me, i could define another output of the function model to be the norm of error_model and try to minimeze this with fminsearch and then use this values for lsqnonlin, i am correct?

if i am, how can i use a different output of the function that is not the first output? if i want to use the same function for fminsearch and lsqnonlin, i could use:

function [norm_error,error_model,h]= dispmodel_2param(param,t,x,y)
...
end

so then i fminsearch will work (from what i understand from you) but then lsqnonlin no (or at least not with the same line of code that was before...)

[param,resnorm,~,exitflag]=lsqnonlin(@(param) dispmodel_2param(param,t,inlet_signal,outlet_signal),param0);

thanks

[Walter Roberson]

        param0=fminsearch(@(param) sum((dispmodel_2param(param,t,inlet_signal,outlet_signal)-outlet_signal).^2), param0);

[franco otaola]

lol, sometimes we dont think in the easier solutions....

thanks a lot for the help!

either way, i still have the doubt, can we call fmin or lsq for a i output that is not the first one of the function?

best regards!

[Walter Roberson]

No, not without using an extra function to capture the extra output and return it instead. It would have to be a true function, not an anonymous function, because of the way that MATLAB deals with additional outputs.

[franco otaola]

oh, okey, thanks!