How to control simulink with fmincon using 2 (two) decision variables

I have a simulink file that I want to control. I want to minimize my objective:

function q = obj(cr,tr) % cr, tr are the variables that I want to be manipulated

 sopt = simset('solver','ode3','SrcWorkspace','base','DstWorkspace','base','SaveFormat','Array');
 [~,~,yo]= sim('sgs2.slx',[0 10], sopt);

 q=abs(yo(end,1)-yo(end,2));

end

I call fmincon

x=fmincon(@obj,xo,[],[],[],[],xmin,xmax,[],options);

the optimization runs but fails.

I think that the way that passing the decision variables is wrong but I have no idea what to try.

Answer 1

[Titus Edelhofer]

Hi,

do you have other variables in your model apart from cr and tr? If not, you should select 'current' for SrcWorkspace and DstWorkspace. This way Simulink would grab the variables cr and tr from the input variables.

Second: your objective function needs to be declared differently. It must have the form obj(x) where x is a vector (of size two in your case). So you would need to change

function q = obj(x)
cr = x(1);
tr = x(2);

Titus

[Stefanos Kalandaridis]

Thanks for the response Titus, I changed the value of SrcWorkspace and DstWorkspace to 'current' and the optimizer does not runs. Changed the objective to:

function q = obj(z)

        cr=z(1);
        tr=z(2);
        sopt = simset('solver','ode3','SrcWorkspace','base','DstWorkspace','base','SaveFormat','Array');
        [~,~,yo]= sim('sgs2.slx',[0 10], sopt);

        q=abs(yo(end,1)-yo(end,2));
end

It runs well but with wrong results.

I'm trying to make the optimizer to try different values for cr and tr so the objective can be minimized. Maybe I'm wrong where the optimizer "communicates" with the simulink file.

[Stefanos Kalandaridis]

Finally figure out what was wrong. I was running an extra simulation and this was confusing the optimizer. Thanks for your help.

By the way do you have any idea why the optimizer tries the same values for the variables three times before the next try ?

[Titus Edelhofer]

Yes, I do: the values are not exactly the same, but very similar. For the jacobian it computes the finite differences

df/dcr ~ (f(cr+h, tr)-f(cr,tr))/h

and

df/dtr ~ (f(cr, tr+h)-f(cr,tr))/h.

So you see three (similar but not equal) calls to f for each iteration. More generally, for n variables you need n+1 function evaluations per iteration (neglecting the fact that the optimizer sometimes updates the jacobian without recomputing it).

Titus

[Stefanos Kalandaridis]

Thank you very much for the information.