Without a minimal working example see the FAQ, more concrete guidance would require reverse engineering the problem, so I'll offer some general guidance.
Every problem in R can be approached with advantage with school algebra in mind: f(x) = y, where the three objects, f, x, y (in R everything is an object) are
x what is at hand
y what is desired
f a function to convert x to y
Each of these objects may be, and in the case of this problem are, composite, consisting of other objects. In particular, f will be analogous to f(g(x).
For the problem
x the [unknown for purposes of the question] object [redes] on 20 tibbles for each of 4 networks [of unknown object type] repeated over 5 network objects
y is a table of results
g performs an operation [unknown for the purposes of this question] on each network using some or all parts of each tibble that extracts the results desired
f collects the return values of g and records them in y.
While it is possible to do this with the for function, procedural/imperative programming style is not R's long suit. For this type of problem, apply, sapply, lapply or its purrr front ends in its map family are preferred.
To proceed without a reprex, compose f and g to deal with the single case of one repetition of one network on one tibble. By running the ANN-related tests and placing them in the receiver object y.
Recall that, since functions are first-class objects, they may be given as arguments to other functions. This has two applications: first, h = f(g(x); second m(h(z)$ will operate over k_i ...k_n, j_i ... k_n where z is some object that encapsulates the indices i and j.