Deep Learning and Civilization

[Bibor]

I'm fascinated by this topic as I work for IBM and we are investing a lot of effort in this area. Will see if we have any tools we can bring to bear on this.

I choose to disbelieve that we have a man (or woman?) working at IBM on artificial intelligence and you still didn't make them create a new AI for Civ4. It's been 10 years! Outrage!


On the topic:

I don't know anything about machine learning, really, so here's a question:
Let's say that someone really creates a machine learning AI for Civ4. Which would mean the human and AI would play on the same difficulty level. Yes, the AI would learn from the numerous games it would play against other AIs (that would cheat, unlike this AI). This would ultimately lead to an AI that plays like a human - optimizes processes, beelines for opportunity windows (Oracle, Cavalry etc.). However, and this is my question really, it would also need to learn to feign. Can it learn to feign? Because, at least my logic, is this: since the human makes so much sacrifices to beat a Deity AI, it makes him vulnerable. if a ML AI would do the same, wouldn't the human be able to exploit this? Imagine the ML AI learning how to beeline Oracle perfectly, just to be bum rushed by a human axe rush, because the human predicted the AI would beeline the Oracle and neglect techs and defenses. I watched enough SC2 matches to see that this happens there all the time (human vs human). Mind games over mind games. Now, the ML AI might be perfect, but what if the human rolls a dice every match to see whether it will try to disbalance/rush/break the ML AI or attempt a beeline of his own? So the ML AI would learn to be more careful. In any case, I don't think Civ4 would be a less complex game than chess or go, *especially* since the map is unknown (or should be, to a ML AI).

 

[aitkensam]

My gut instinct (as an experienced Go player) is that Civ IV would be harder for AI than Go or Chess.

Civ takes more moves than chess and a fairly similar amount to Go, But Civ will branch very quickly compared to either go or chess. Whereas in chess or go the number of possible moves stays fairly steady (~20-50 or so in chess vs 200-400 in go), in civ it will continue to increase as the game progresses and be thousands upon thousands. By late game I imagine brute force won't hold out much hope...

But my main worry for the AI is with regards to how it would evaluate its position in the early game, and hence how it would be able to make sensible decisions.

However, I am sure some clever AI guy could come along and show me how naive I am.

Slightly off topic, the DeepMind team are planning to tackle Starcraft 2. I've never played it. Would it require similar skills to Civ IV for an AI?