How strong AI do you like to play against?

[Alpakinator]

I'm meddling with CBP AI and started wondering if others like to play against very competent AI. Hard to answer, I know, since we still don't have truly good AI in any civ game to compare against, and CBP AI is the best . But:

If CBP AI devs improved it, such that it:

• never builds redundant units
• remembers enemy unit positions and takes it into account,
• hides it's true intentions perfectly,
• targets best wonders every time etc.

would you appreciate that? I'm just thinking, if there are AI imperfections that are good for gameplay. Leader specific preferences seem like something that should be reasonably maintained at all times even if it makes AI weaker.

Cause you know, if we took it to the extreme and had a Civ 5 neural network, then it'd be better at winning than any player ever, by a long shot. Would it be fun if it was present only in deity difficulty for example, with some appropriate handicaps?

 

[CppMaster]

It would be cool to Play again neural network AI that is very good and then just się difficulty as we have now to handicap it or the player. Like current Immortal would be as hard with neural network AI Chieftan

 

[andersw]

So hard to say, on some VP versions AI have blanketed the land with units easy popping a handful units from what seems like nowhere.
AI also have had much bigger combat bonuses.
Theres been several versions where AI either sacrifice units to get kills or retreats to save units.
It is not easy to say what way is best way to make an AI strong.

 

[TheOrpheus]

If the AI gets any better I guess I'll just have to play on Cheiftain. VP has gotten more and more difficult over the years. I find myself having to play at Prince if I want to do anything other than conquest with a militaristic civ. I suppose improvements can be made to decision-making and strategising that shouldn't outright make the game much more difficult. But I'm struggling a lot on harder difficulty these days.

 

[Alpakinator]

Yea, it's important to keep the same difficulty spectrum. After improving AI, it can be nerfed more on lower difficulties. From what I understand, current AI on lower difficulties intentionally chooses suboptimally, on top of having smaller buffs to yields.

If the AI gets any better I guess I'll just have to play on Cheiftain. VP has gotten more and more difficult over the years. I find myself having to play at Prince if I want to do anything other than conquest with a militaristic civ. I suppose improvements can be made to decision-making and strategising that shouldn't outright make the game much more difficult. But I'm struggling a lot on harder difficulty these days.

 

[TheOrpheus]

Yea, it's important to keep the same difficulty spectrum. After improving AI, it can be nerfed more on lower difficulties. From what I understand, current AI on lower difficulties intentionally chooses suboptimally, on top of having smaller buffs to yields.

I suppose that's a testament to just how improved the AI is in general, that even with suboptimal choices, they can be surprisingly hard to deal with. Credit to the VP team. I just haven't been able to keep up I guess (used to play Emperor/Immortal).
I consistently play on Prince now. But I don't play OP strategies or try to "game" the game, I just pick a Civ and try to play a game that I think fits with their traits. I've had some easy domination wins on King though. Just depends on how well you get going early on.

 

[Alpakinator]

I suppose that's a testament to just how improved the AI is in general, that even with suboptimal choices, they can be surprisingly hard to deal with. Credit to the VP team. I just haven't been able to keep up I guess (used to play Emperor/Immortal).
I consistently play on Prince now. But I don't play OP strategies or try to "game" the game, I just pick a Civ and try to play a game that I think fits with their traits. I've had some easy domination wins on King though. Just depends on how well you get going early on.

Playing with certain more or less OP strategies, especially with specific map setting in practice lowers difficulty by 1-3 levels, so dw.
I think improving combat AI is the priority right now, since for most players, winning domination is the easiest, while science or culture victory with no conquest is much harder in comparison.

 

[andersw]

If the AI gets any better I guess I'll just have to play on Cheiftain. VP has gotten more and more difficult over the years. I find myself having to play at Prince if I want to do anything other than conquest with a militaristic civ. I suppose improvements can be made to decision-making and strategising that shouldn't outright make the game much more difficult. But I'm struggling a lot on harder difficulty these days.

I misread op as "strong in wars" but strong can ofc be applied to any type of game play.
I probably should try some more peaceful play but I like the micromanagement or wars, tile improvements etc.
Humans can get a lot more yields from wars because of better positioning and choice of units while peaceful play is possibly more even?

 

[DizzKneeLand33]

Here, Emperor, plus or minus one level (in vanilla, Immortal is the challenge).

 

[msw1]

Currently in every civ game human players are better than AI players ever will be, so the AI needs a lot of "cheats". I think it would be fun to swap roles with the AI, if such a thing were possible, i.e. play against an AI that is far better than you but give yourself some kind of very strong starting advantage. I do something kind of like this sometimes by setting up the map settings to be perfect for my civ but then set the difficulty to Deity.

 

[Recursive]

Currently in every civ game human players are better than AI players ever will be, so the AI needs a lot of "cheats". I think it would be fun to swap roles with the AI, if such a thing were possible, i.e. play against an AI that is far better than you but give yourself some kind of very strong starting advantage. I do something kind of like this sometimes by setting up the map settings to be perfect for my civ but then set the difficulty to Deity.

I gave human players the ability to use all of the AI's difficulty bonuses, including the new columns added by Milae in his alternative difficulty mod. See the very long comment at the end of MODS\(2) Vox Populi\Database Changes\Difficulty\DifficultyChanges.xml.

 

[Tekamthi]

I'm under the impression that nearly all the possible ai intelligence has been just about wrung out of the existing system. I play immortal most often cuz it keeps my success rate slightly under 50%, so for me the AI is competent enough already. My success rate on deity has dropped to maybe 5%, too narrow to enjoy. I'll sometimes play below immortal to try out a new and unconventional strategy, before trialing it on immortal.

Cause you know, if we took it to the extreme and had a Civ 5 neural network, then it'd be better at winning than any player ever, by a long shot

Maybe y'all are familiar with aspects of the mod I am not, but wouldn't we need some kind of REST API first? No such effort exists afaik, and building such is not necessarily trivial. What exists in lua is all sandboxed afaik. I think it might be possible however? Anyway it seems a bit of a cart-ahead-of-the-horse thing to make plans for neural network when no support exists for such. I like the idea tho fwiw, as it might allow for more complex and life-like systems

 

[Alpakinator]

I'm under the impression that nearly all the possible ai intelligence has been just about wrung out of the existing system. I play immortal most often cuz it keeps my success rate slightly under 50%, so for me the AI is competent enough already. My success rate on deity has dropped to maybe 5%, too narrow to enjoy. I'll sometimes play below immortal to try out a new and unconventional strategy, before trialing it on immortal.



Maybe y'all are familiar with aspects of the mod I am not, but wouldn't we need some kind of REST API first? No such effort exists afaik, and building such is not necessarily trivial. What exists in lua is all sandboxed afaik. I think it might be possible however? Anyway it seems a bit of a cart-ahead-of-the-horse thing to make plans for neural network when no support exists for such. I like the idea tho fwiw, as it might allow for more complex and life-like systems

That neural network idea was just to illustrate a potentially too strong to be fun AI, as an example. I think it's not viable for CBP. Just like with AphaGo or LLMs It'd take millions of $ to train it. It'd need to complete millions of civ V games with itself to learn.

 

[Tekamthi]

It'd take millions of $ to train it. It'd need to complete millions of civ V games with itself to learn.

hypothetically, there are millions of dollars worth of gpu's running civ 5 every day -- someone maybe could figure out a way to crowd source it....

but pushing for an api would be more immediate and provide other benefits possibly, thats all im saying.

 

[Alpakinator]

hypothetically, there are millions of dollars worth of gpu's running civ 5 every day -- someone maybe could figure out a way to crowd source it....

but pushing for an api would be more immediate and provide other benefits possibly, thats all im saying.

I don't think I understand. Do you mean using an AI that some company already trained, that we'd plug to civ 5 via an API?

 

[Tekamthi]

I don't think I understand. Do you mean using an AI that some company already trained, that we'd plug to civ 5 via an API?

in one form the training data is being generated by all the ppl playing civ 5 -- afaik there is no way to extract sufficient detail from the game engine but some kind of REST API would allow the training data to be extracted, in theory -- the someone's gotta figure out how to have one of these learning algorithms assess it all. seems possible to me as armchair AI expert lol.

but without API there's no way at all to get any kind of training data out, unless one of these models can figure out how to read the visual UI off the screen itself -- think we're a long way off from that yet

 

[Alpakinator]

in one form the training data is being generated by all the ppl playing civ 5 -- afaik there is no way to extract sufficient detail from the game engine but some kind of REST API would allow the training data to be extracted, in theory -- the someone's gotta figure out how to have one of these learning algorithms assess it all. seems possible to me as armchair AI expert lol.

but without API there's no way at all to get any kind of training data out, unless one of these models can figure out how to read the visual UI off the screen itself -- think we're a long way off from that yet

Ah, human training data. That's an interesting idea, cause as you say, we could get it for free if mod users would agree.

 

[Tekamthi]

i recently uploaded the vp repo to chatgpt o3-mini and it was reasonably good at doing basic modding tasks -- i had to inform it of a lot of things but it did write useable lua with some non-trivial logic involved in the end

in that sense maybe it could help write lua-based AI code for specific things, new gameplay features etc. havent tested it this far yet, but if you check out the 'maritime weather+' mod on the forums here, the code for spawning and moving storms was all written by ai, according to my (detailed) design specifications (and revision instructions)

 

[Alpakinator]

Still, there aren't enough CBP players to provide enough data. I wonder how much $ of compute would it cost to train a model on logs from let's say 1 million full civ 5 games.

For context The first AlphaGo learned from 30 mln games that were a mix of games with itself, and records of human games. Later, AlphaGo Zero didn't use any human data, and only needed 4.9mln games to become even better than first AphaGo, because humans are bad at go.

 

[Alpakinator]

i recently uploaded the vp repo to chatgpt o3-mini and it was reasonably good at doing basic modding tasks -- i had to inform it of a lot of things but it did write useable lua with some non-trivial logic involved in the end

in that sense maybe it could help write lua-based AI code for specific things, new gameplay features etc. havent tested it this far yet, but if you check out the 'maritime weather+' mod on the forums here, the code for spawning and moving storms was all written by ai, according to my (detailed) design specifications (and revision instructions)

Cool mod! Claude 3.5 is pretty good at undersanding the repo too. Very helpful cause it's huge and complicated.

Btw. There's still lots of intelligence to be extracted from current CBP AI. EG: It fails at naval warfare, cause it doesn't build enough ships. I'm trying to fix that.

While you can build your 30 ship navy in early reneissance, the top snowballing AIs, that are 7 techs ahead of you, have 2x more cities and most wonders, will not build more than 10 ships. Literally every tile of their islands will be covered with tercios, knights and crossbowmen. So you with your 30 galleases and few caravels can just safely hit'n run their units and cities and capture everything.