Fitting a function to data (fminsearch) with limits

Hey!

I try to fit a model to measured data. I already got a real close with the fminsearch option. However, I know the model will never fit completly to the measured data. I know the range wherein the unknown factors should be, how can I include this in the function.

This is a part of my code;

p= [0.31;114;3.5];
y2=fitfunc(TempZone1day,TempZone2day,TempZone3day,TempZone4day,TempZone5day,buitenTempday,FlowZone1day,FlowZone2day,FlowZone3day,FlowZone4day,KNMIwindday,SMA,p);

figure(1)
plot(tijd,WarmteCvday,'+',tijd,y2,'-')

a0=p;
aBest = fminsearch(@(a) SumErrfun1(a,TempZone1day,TempZone2day,TempZone3day,TempZone4day,TempZone5day,buitenTempday,FlowZone1day,FlowZone2day,FlowZone3day,FlowZone4day,KNMIwindday,SMA,WarmteCvday,tijd),a0);
disp(aBest)

For example, I know that p(1) should be within 0.1-0.5, p(2) within 100-200 and p(3) within 2-6. Because now aBest (the best combination) is [-0.0328; 61.8202; 0.4375], which is not even a possible option. How can I include these ranges?

Answer 1

[John D'Errico]

fminsearch has no capability to take bounds on the search. If the objective is such that a better result lies outside of where you want it, too bad. :)

Having said that, you can use fminsearchbnd , a tool found on the file exchange. It does allow bounds on the variables. Just download and install that tool on your search path, then use it instead.

http://www.mathworks.com/matlabcentral/fileexchange/8277-fminsearchbnd--fminsearchcon

[loes visser]

Super, thanks! It's working :)

Answer 2

[Jie Jian]

Or you can use the function 'mapping_parameters.m' to transfer unbounded parameters to bounded ones

https://au.mathworks.com/matlabcentral/fileexchange/73872-mapping-unbounded-parameter-to-bounded-one-for-fminsearch

Answer 3

[kursat cihan]

John D'Errico....much love from Germany, helped me a lot!!!

Bachelor Thesis in Material Modelling, used it for a parameter optimization in bringing simulations together with experimental data...PEACE

[Delme]

Haha, i'm doing exactly the same in my thesis!

Answer 4

[Stefan Schuberth]

you can use q=f1*atan(p)+f2 to construct a limited parameter q from an unlimited parameter p :)