I am conducting simulations for Genetic Algorithm using optimtool('ga') of Matlab.
I would like to minimize the difference of two arrays so that I can fix the value
of a parameter which is the population.
The first array includes some measurement values when the second includes the theoretical values - the unknown values. Their difference is the objective function.
So I would like to minimize an array of differences. Is this possible to obtain?
I ask this because until now matlab brings me the following error:
Error using makeState (line 56)
Your fitness function must return a scalar value.
Do you have any idea? Thank you in advance.
Hi Panagiotis,
As suggested, you can minimize the sum of the squares of the differences instead to get a scalar-valued target function. MATLAB also has some built-in functions like 'lsqnonlin' that form this sum internally, so you can use matrix-valued functions as well. I have found 'lsqnonlin' efficient and flexible in various uses.
Regardless of whether you compute the sum yourself or input the matrix, it is good to think carefully about the weighting of the differences in the sum, because this has a drastic effect on the found optimal point if your system is overdetermined. Note that you can multiply each difference with a weight factor (having in mind that essentially only the relative weights matter). This is particularly important if the differences are for completely different variables (with different units and characteristic magnitudes). Then it is good to think a way of making them comparable to one another (and dimensionless), so that the differences become weighted properly and justly in the optimization. One way is to consider the relative errors, i.e., divide each difference by the target/measurement value, and then sum the squares of these.
use MSE function of all error differences you will get a scale value.
Hi Panagiotis,
As suggested, you can minimize the sum of the squares of the differences instead to get a scalar-valued target function. MATLAB also has some built-in functions like 'lsqnonlin' that form this sum internally, so you can use matrix-valued functions as well. I have found 'lsqnonlin' efficient and flexible in various uses.
Regardless of whether you compute the sum yourself or input the matrix, it is good to think carefully about the weighting of the differences in the sum, because this has a drastic effect on the found optimal point if your system is overdetermined. Note that you can multiply each difference with a weight factor (having in mind that essentially only the relative weights matter). This is particularly important if the differences are for completely different variables (with different units and characteristic magnitudes). Then it is good to think a way of making them comparable to one another (and dimensionless), so that the differences become weighted properly and justly in the optimization. One way is to consider the relative errors, i.e., divide each difference by the target/measurement value, and then sum the squares of these.
yes ofcours you could make the difference in simulink e=u1-u2
save e in file istead (to workspace)
and your f(t) = trapz(e.^2)
- Badges
- Science topic