Never got any reply to this link although it's the third time I mention it in three different posts over the last months!
Ok, let me reply myself with a question. Is this too disturbing, or underwhelming, or off-topic, or hard evidence that yes, machine-learning is possible and affordable for Civ6? Any expert opinion here?
I think that I read the article back when it came out, but didn't remember some key details. So i re-read it today. In my day job, I work in a department that supports teams doing AI, but I don't work in it myself. I played a handful of games in FreeCiv more than 15 years ago.
1. In the accompanying video, the ML logic first settles cities, constructs improvements, and builds its empire. Later in the video, it is shown defending itself from an invasion. It pursues a space race victory, since that is one of the two victory types in Civ2.
2. In the text, the learning aspect is described as using machine reasoning, where experience from humans is entered as words and phrases. Put another way, the company could have fed in one of the CivFanatics strategy guides recommending techniques in the early game, combat, and achieving victory.
3. FreeCiv, like Civ2 and Civ3, allows unit stacking and traversing over mountains. Mountains became impassable in Civ4, and remained impassable (with limited exceptions) in Civ5 and Civ6. Thus, the logic for moving units is much less complex than Civ6.
4. The article makes no reference to the computing resources needed to run their ML logic while it plays the game.
So, I believe it would be possible for Arago (the company) to train HIRO (their software) to play Civ6. Since they would ingest human-generated experience about gameplay, such as how to use siege weapons to assault a city, I would expect HIRO to do that competently. I would expect that HIRO would pursue a space victory in Civ6 competently. HIRO might even pursue a culture victory effectively.
Key questions:
- What would be the economic model for making money off such an AI? Perhaps online play against HIRO, where a server in the cloud is running the logic? Would players pay a subscription fee to have a HIRO AI opponent? Would it be profitable for Firaxis to include a surcharge in each game purchase to have that AI available? Given what we've read about the number of hardcore fans and the number of casual players, I'm not sure how profitable it would be to charge all purchasers for the HIRO AI?
- What would be the impact on fun for the human players by using such an AI as an opponent? This logic would tend to pursue a path it believes to be optimal to pursue certain victory conditions. Might its metanarrative "always go for space", "always go for XXX" become predictable? How would difficulty levels be implemented? What happens when 2, 3, 4 HIRO AIs all start on the same landmass?
- If the logic is implemented on the player's PC, what are the HW requirements for that PC? How would that be different from the current GPU requirements for gaming PCs today?