What’s new in 8 years?

[Arcaian]

Chat GPT is for talking, not playing games.
AlphaGo decisively won over all Go players in 2015.
And AlphaStar was running GM on all 3 races (3 of the best 200 players) at Starcraft 2 in 2019.

Civ exists between those two as far as difficulty goes and since it's a game (not a language), unsupervised learning or learning from a small dataset is doable. The real issue is cost. These things aren't plug and play. You need to pay for lots of compute power and smart people to get it there.
While video game business is all about cutting margins and cheap construction/factory worker-level engineering.

IMO, it's pretty ridiculous to say that optimizing Civ AI is less challenging than Starcraft 2 AI. A game of Starcraft is over in 15 minutes, a game of civ takes at least several hours, for most speed settings into double-digit hours I'd suspect. If you want to play optimally, that decision on turn 1 to start with a scout or start with a slinger is incredibly relevant on turn 150. That's a nightmare to code efficient AI for, especially if you're using ML-based methods which simply have to learn relationships in the data instead of having the logic coded out for them. Optimal play in Civ involves seeing something on turn 10 with your scout, realising that it's going to be useful many turns down the track - maybe a lovely spot for an industrial district with an aqueduct or a dam, a defensive chokehold when you've got an aggressive neighbour, or even a lovely city to chop out a spaceport for a science victory - and acting on that now. That's one of the hardest things to get an AI to do - if you're using logic-based methods, you have to try and get the AI to make some logical inference about that without explicitly calculating out the 50/100/200 turns in-between discovering the location and getting the benefit from it, because that would ruin turn times. If you're using ML-based methods, you somehow have to get the method to understand that it was this one decision at turn 24 that led to the benefit at turn 124, which requires a frankly ridiculous amount of computation and data.

 

[UWHabs]

I don’t think machine learned AI is a good idea for Civ. This suggestion assumes that a hyper competitive AI is what’s best for the series, and I think that’s not correct.

I certainly don’t want AI to play like someone who takes multiplayer Civ competitively. Then you have an AI that will adhere to some unfun cheesy “meta” strategy focused on min-maxing.

I want Civ AI to be immersive and make me feel like I’m in a real alternate world. If I wanted to experience annoying human min-max tactics, I’d play multiplayer with random people.

It would take them probably 20 minutes to create a Deity-level AI that can always win against the human:
Have them immediately declare war against the human player, immediately build as many units as they can, and rush the human player, and I'm sure they'd have like a 90% win rate against even the best human players. A turn 15 rush of 4-5 suicide warriors against a human city you'd have to have a seriously great defensive position to be able to hold that.

But that AI would not be fun to play against. And would probably fail miserably on lower levels where the AI doesn't start with all their extras (although I'm sure you could just code them to build up a big military and rush the human early, since most human prince-level players are not rushing out walls for defense asap).

You want an AI that if they have a ton of land nearby, turns into an expansionist civ. You want one that you can possibly become friends with given certain conditions (maybe based on their internal agendas). You want AI that plays to their civ's abilities (ie. you want Khmer to be a tall civ that leans on holy sites, you want Qin to be an ancient era wonder builder, etc...).

The civ 6 AI isn't terrible at that. They could be better, yeah, for sure. And it's still frustrating when you like have this big epic game going, and then your big warmonger neighbour only has like 4 units from an era ago, and they constantly move them into terrible spots, don't shoot against your units, etc... and you end up with an easy time taking them out. So yeah, I do hope that they can get a little more into that for the next version, to at least be able to give you a challenge. Or even better, if you could play at Deity without the AI starting with the absolutely insane bonuses that they do. But yeah, trying to tune a massive learning algorithm for a game like civ isn't really reasonable.

 

[Leyrann]

It would take them probably 20 minutes to create a Deity-level AI that can always win against the human:
Have them immediately declare war against the human player, immediately build as many units as they can, and rush the human player, and I'm sure they'd have like a 90% win rate against even the best human players. A turn 15 rush of 4-5 suicide warriors against a human city you'd have to have a seriously great defensive position to be able to hold that.

That would work... for one day.

And then the humans would adapt, start by building two or three warriors, and hold by using their defensive advantages. Meanwhile, because they're building a few fewer warriors, they get to build more of other, useful things and get ahead.

And suddenly, you're back to the AI requiring bonuses to be a challenge. Except now it's more annoying to pay against because it tries to cheese you.

But that AI would not be fun to play against. And would probably fail miserably on lower levels where the AI doesn't start with all their extras (although I'm sure you could just code them to build up a big military and rush the human early, since most human prince-level players are not rushing out walls for defense asap).

Oh wait, you were imagining that they kept their Civ VI Deity level bonuses.

The thing with bonuses is, you start by building an AI, and then you apply bonuses based on the strength of that AI to get the correct difficulty. Not the other way around.

 

[vorlon_mi]

It would take them probably 20 minutes to create a Deity-level AI that can always win against the human:
Have them immediately declare war against the human player, immediately build as many units as they can, and rush the human player, and I'm sure they'd have like a 90% win rate against even the best human players. A turn 15 rush of 4-5 suicide warriors against a human city you'd have to have a seriously great defensive position to be able to hold that.

But that AI would not be fun to play against. And would probably fail miserably on lower levels where the AI doesn't start with all their extras (although I'm sure you could just code them to build up a big military and rush the human early, since most human prince-level players are not rushing out walls for defense asap).

This.

It would be straightforward to code a computer player logic to follow the "Always War" variant that some players use in other games in the franchise. IIRC, it might have even been a toggle option in Civ4. That is, declare war upon meeting any civ; build units and military buildings;attack. Find the weakest nearby civ and smash it. Repeat often enough, and the AI factions will certainly overwhelm most human players, especially with Deity-level starting advantages.

But that AI would NOT be fun to play against. The sales for a game that included that AI would be less than the sales for a game where the AI uses more of the game features and has more of the traits that all the franchise games have.

Having said that, I do want certain civs to be more aggressive and expansionist. I want them to successfully invade and take cities from their AI neighbors. In my experience, the current AI logic -- in Civ6, in BERT, in Civ3, other franchise games -- can successfully work towards victory conditions that require *building* something: a victory wonder, accumulating diplo points, winning a vote in the UN, sending lots of rock bands to generate tourism. They have much less success in achieving either the Civ2/Civ3/Civ4 definition of military victory or the Civ5/BERT/Civ6 definition.