How can I set the numbers of decimals (precision) of the solution of fminunc?

I am minimizing a multivariable function using fminunc and I am interested in a solution with just two decimals (ex. A=3.04, not A=3.0421). This is because I am minimizing the error through calibrating some coefficients and I want those coefficients to have a certain precision.

[Birdman]

Check my answer.

[Patricia Mares Nasarre]

I have just read it and I am trying to see how to implement it in my code, so I can give feedback.

minimize=@(CG)rMSE_hc2est(CG, manto, Hs, Tm, Rc, hc2exp);
x0=[0.1 0.1 0.1 0.1 0.1];
[CG, rMSE_obtained]=fminunc(minimize, x0, options);

(CG is a vector)

[Birdman]

Which value you want its decimals to be removed in this code?

[Patricia Mares Nasarre]

Values in CG vector

[Birdman]

Ok same thing:

CG=vpa(CG,3);

at the end of your code

[Patricia Mares Nasarre]

It does not work: it transforms the CG vector in a symbolic matrix with no numbers

Answer 1

[KL]

Use the StepTolerance option,

 options = optimoptions(@fminunc,'StepTolerance',0.01);

https://www.mathworks.com/help/optim/ug/fminunc.html#input_argument_d119e59170

also https://www.mathworks.com/help/optim/ug/tolerances-and-stopping-criteria.html

[Patricia Mares Nasarre]

I have already tried it and I obtain the following error: Error using optimoptions (line 124) Unmatched parameter name 'StepTolerance ' must be a string scalar or character vector that can represent a field name. I have also tried to use 'TolX' with optimset, but the result is still showing so many decimals.

Answer 2

[Alan Weiss]

It is very difficult to know how the option settings affect the resulting solution accuracy. The only solver I know that is reasonably straightforward this way is patternsearch, and even that solver can be fooled.

My suggestion is simply to leave the options at their default values, if the process is not taking excessive time. If you want answers much faster than you currently are getting, first I would try to give better initial points x0, and only afterward gradually relax the step tolerance or optimality tolerance. But be careful, relaxed tolerances can have unwanted consequences, such as convergence to a point that is very far from optimal.

Alan Weiss

MATLAB mathematical toolbox documentation

[Patricia Mares Nasarre]

I guess I haven't explained properly my problem. I want the solver to "jump" from 0.53 to 0.54 between iterations (minimum step), so the solution has 2 decimals and it doesn't look like: parameter=0.5321

[Alan Weiss]

I think that I still might not really understand what you want. If it is a display issue, you can choose to show only those decimal values that you like. See format or fprintf. If you want to control the solver iterations themselves, then I suggest again that you not try to do that, unless the solver is taking way too long.

Alan Weiss

MATLAB mathematical toolbox documentation

[Patricia Mares Nasarre]

It is not about the speed of the computational speed or a display issue. I want the minimum error (function to minimize) with coefficients with just two decimals

Answer 3

[John D'Errico]

"I want the solver to "jump" from 0.53 to 0.54 between iterations (minimum step), so the solution has 2 decimals and it doesn't look like: parameter=0.5321"

That is a discrete problem. Essentially, you want the solver to look at only a set of integer values, and fminunc is NOT such a solver. (I know, they are not integers that you want. Multiply them by 100, and they would be integers.)

The point is, you cannot use fminunc to solve that problem.

You CAN use tools that can handle optimization over an integer set. I recall GA is such a solver. intlinprog also allows a partially integer search space, if linear programming is an option.

[Patricia Mares Nasarre]

I will try to use integers with the ga solver, but there must be a way to control de step and shape of the parameters to optimize.

Modified: I am using ga function and the results of the minimization are awful

[John D'Errico]

If awful, that may just indicate you don't know how to use the tool properly.

As far as controlling the "step" of the parameters. Require integers, then divide by 100. WTP?

[Patricia Mares Nasarre]

The relationship with the coefficients is not linear, so it is not that straight forward.

I would also like to obtain results with diferent precisions, such as values ended in 0 and 5 (0.45, 0.50...)

Answer 4

[Romina Jahangiri]

I have same problem. I want to get 2 digits after decimal point in the result of fmincon. Did you solve your problem?