For this or other successful genetic algorithms, it would seem clear you still can’t know if it’s an optimal design simply because in most cases the number of designs tested would be a small fraction of the possible designs.

However, things like useful sound recognition being done with only a small number of logic gates (commenter below provided a nice article, thank you) make it hard to imagine doing much better.

I wonder if the process can be shown theoretically to offer any help in guaranteeing minimum bounds w.r.t. the optimal case, even if can’t be fully proven to be optimal.