Is it possible to create an optimizableVariable of integer type as an array with step different from 1?

Hello everyone!

I'm trying to perform a Bayesian optimization to find the best hyperparameters for a deep neural network and I want the variable MiniBatchSize to just have, in this case, three possible values: 8, 10, 12. I thought I could declare optVar as if it was an array

interval = 8:2:12;
optVar = optimizableVariable('MiniBatchSize', interval, 'Type', 'integer');

This code raises the following error:

Error using optimizableVariable/checkRange

'Range' value must be a string array, a cell array of character vectors, or a numeric vector of length 2.

Am I missing something? Is there another way to perform this?

Thanks in advance.

Answer 1

[John D'Errico]

Suppose X is an integer variable, that takes on only the values 1:5.

But what you want are non-integer levels of some parameter. For example, perhaps you want levels at [1.5, 3.5, 5.5, 7.5, 9.5]?

Inside your code, just write

Xhat = 2*X - 0.5;

Do you agree that Xhat takes on the desired levels?

They need not even be at all regular. Suppose you want levels at [2 3 5 7 11]? Inside your code, you might do this:

levels = [2 3 5 7 11];
Xhat = levels(X);

Clearly I could have chosen ANY levels there.

In your specific case, X is an integer variable that takes on the values 1:3. But you want it to take on the values [8 10 12]. Again, inside your code, write this:

Xhat = X*2 + 6;

Do you see that Xhat takes on only the values [8, 10, 12]?

[Jon Gutiérrez Zaballa]

Yes, I see it. I just thought it was not possible to modify the scope of the optimizableVariable inside my code. Let's see if it works. Thank you very much.

[John D'Errico]

It is your code. It can do anything you want, as long as you return an objective.

Answer 2

[Walter Roberson]

Optimization variable 1 to 3 and multiply by 2 and add 6.

[Jon Gutiérrez Zaballa]

Thanks but I don't get your point. What do you mean? Could you please write the piece of code to do that?

[Walter Roberson]

Ah, sorry, I didn't realize that the Stats toolbox had its own kind of variable creation; the instructions I gave were for Problem-Based Optimization

[Jon Gutiérrez Zaballa]

After rereading again your answer I've finally got your point and I think it might actually work. Thank you very much.