Good question!. In my opinion man made procedures are based on modelling thinking mainly on:
1.- correct unwrapping
2.- customer hardware (mainly mid or low range) (low polygonization, LODs, precalculated light)
3.- # nodes for lightmassing calculations
I work on distributed swarm agents & coordinator for lightmassing and it’s very common see iddle nodes while others are still computing at 50% 60% total calculations due bad distributed calculations ( big meshes vs. little meshes balance)
I don’t know how algorithms could do the same without very deep learning…
Here you can see what I’m talking about (sorry it’s in spanish)