> I thought the routines that handle the attract modes already did this sort of thing.
Unlikely. Attract mode AI is always either playback of canned input or very, very simplistic. Especially on DK ROM space was super expensive (and it's not like they're going to run a neural net on a Z80).