Since all the NNs have the same architecture, the breeding is done by taking some parameters from one parent NN and some parameters from the other to populate the weights of the child NN (in this case, indices that are 1 mod 2 are taken from the first parent, the others from the second one). In addition to this there's some random tweaks of some parameters (20% chance that any given parameter will be changed, with the change being uniformly random from the range [-0.1,0.1]).

Thanks :)

Yup, I (as a developer of this thing) haven't created a documentation yet, but it's mostly due to the fact, that I have deployed this just yesterday and there are still many things that I'd like to change in it. But I will work on that :)

Edit: The other commenter actually bothered to look at the code. Scratch my comment.