That is my favorite part of GA. Gradient free optimization but it turns out making a good fitness function is hard and like 70% of the time it just exploits some assumptions or gap you have in your theories. Really reveals the problem in different ways that traditional ML.