Thanks for the pointer, but I'm afraid I have zero knowledge of ML and wouldn't be any help.

I have to say I'm a bit sceptical that this approach can work purely based on win/loss conditions - an attacker victory in particular can often take hundreds of moves in Tafl games.