Guys, let's stop talking about ML, it is not that easy and simple to use ML in gaming industry. I know and I introduced one example of SC, but that kind of activity would take some significant amount of computational resources, and some degree of man hour from experienced researchers. Also, such feature requires huge amount of big data, but this game don't even have that. Those would be impossible for this studio and probably any other similar game studios. Like I mentioned, someday 'holy grail of wargame' would come with such AI, but I'm not sure when will be the day.
So... let's forget about it in this post and let's keep some meaningful discussion. I agree with Werezak, we need to looking for concepts which would be easy and straightforward to implemented in the coding, yet good enough to improve AI's behavior.
I understand there is a long way to go even with the resources of the Pentagon, but DARPA is leaving no stone un-turned. This is an interesting article on how they are treating battlefield positions as objects whereby the machine is taught to distinguish good and bad positions and how to improve them based on resources and maneuver. This is not in AlphaZero's wheelhouse currently.
