What kind of neural network do you need?

Question

Neural networks

Suppose there is a certain, alien, black box into which variables are entered and it produces a result, for example: there is a formula ( (x * y) / z), but I don’t know it, but I know that with x=5, y=6, z=2 the result will be 2, at 10, 3, 4 the result will be 7.5 and so on, I collected a lot of statistics (x, y, z| r), but I don’t know what kind of neural network I need and where to start after the general theory, Perhaps you can tell me examples of solving such a problem or in which direction to dig?

Answer 1

Regression analysis is better .
First plot a large number of experiments. By appearance, try to predict what kind of function is hidden there: the simplest linear, quadratic, polynomial.
Then build a model and adjust its parameters so that the resulting function is minimally "erroneous" relative to the experimental data from the black box.

Answer 2

I already answered in the comment to the previous question.
Example: 4,5,9 and 3,6,11
Answer: 3.5, 5.5, 10
Matrix:

3, 3.5, 4 -> -0.5, 0, +0.5
5, 5.5, 6 -> -0.5, 0, +0.5
9, 10, 11 -> -1, 0, +1

It remains to arrange the coefficients in the formula.

Answer 3

A neural network is just a way of approximating data.
You have a large and complex formula (represented in the form of NN), and in it, using the methods of "learning NN", you just adjust the necessary parameters.
That. NS or some other formula is unimportant. The accuracy of approximation to the black box is important.
But why are NNs winning now - yes, all because they have thousands and millions of parameters that need to be adjusted - that is. they are quite flexible. But there is no question of any intelligence here.