Civ 7 using Generative AI

UncivilizedGuy

The Village Idiot
Joined
May 24, 2012
Messages
961
Location
Land of the Lost
AI has always been a sore spot in Civ games. This doesn't have to be the case anymore. Generative AI can be the solution we have been looking for. As a matter of fact I won't be surprised if this is the future of all games. Gen AI can essentially fill a seat in a multiplayer game. The server for the AI could either be provided by Firaxis or another company like Steam. The AI doesn't even have to be Civ games specific. I asked ChatGPT if it knew how to play Civ and it said "of course". Unfortunately there is no way to directly link the server to the game, at least that I'm aware of. Thoughts?
 
i've wondered if it's feasible for a game like civ have a client / server architecture so anyone could develop an AI client. I haven't worked with Generative AI but I would think that would solve the problem of directly linking the server like you mentioned.
 
AI has always been a sore spot in Civ games. This doesn't have to be the case anymore. Generative AI can be the solution we have been looking for. As a matter of fact I won't be surprised if this is the future of all games. Gen AI can essentially fill a seat in a multiplayer game. The server for the AI could either be provided by Firaxis or another company like Steam. The AI doesn't even have to be Civ games specific. I asked ChatGPT if it knew how to play Civ and it said "of course". Unfortunately there is no way to directly link the server to the game, at least that I'm aware of. Thoughts?
Ask ChatGPT how much 34+23+78+124+92+64+75+14+87 is.

What exactly is an AI language model going to do for Civ at this stage? Neither is the AI sufficiently advanced, nor is Civ VII in such an early stage of development that it could develop a game around an AI with capabilities that do not yet exist.

What am I missing here?

I asked ChatGPT if it knew how to play Civ and it said "of course".
Lol. ChatGPT does that a lot. It's remarkably confident about lots of things, including stuff it's completely inadequate to perform.
 
Ask ChatGPT how much 34+23+78+124+92+64+75+14+87 is.

What exactly is an AI language model going to do for Civ at this stage? Neither is the AI sufficiently advanced, nor is Civ VII in such an early stage of development that it could develop a game around an AI with capabilities that do not yet exist.

What am I missing here?


Lol. ChatGPT does that a lot. It's remarkably confident about lots of things, including stuff it's completely inadequate to perform.
All true! How about an AI that is specifically trained on gaming? AI has been proficient at chess for a couple of decades. We may be at the point now where AI can play more complex strategy games like Civ.
 
AI, yes. Chat GPT, not so much:

 
AI, yes. Chat GPT, not so much:

Exactly the example I was thinking of. ChatGPT can't properly play chess, so unfortunately not Civilization either. Though with the right programming perhaps some machine learning could be used.
 
All true! How about an AI that is specifically trained on gaming? AI has been proficient at chess for a couple of decades. We may be at the point now where AI can play more complex strategy games like Civ.
I think it's safe to assume AI will keep on improving and this will be reflected in gaming AI as well.

A few years back I was hoping it would advance quick enough that we could even see some applications in Civ VII. I was probably overly optimistic.

So I'm pushing it a few more years down the road.

I do look up to the day when the discussion on these boards regarding AI will turn from "AI sucks. How can it be improved?" to "AI sucks, how can we make it worse?" or "AI sucks, I can tell when it's self-sabotaging, and it is breaking immersion".

So, bottom line, the discussion regarding what constitutes "good AI" will probably continue even after the AI is capable of demolishing even the best Civ players.
 
I can imagine that some learning model could be constructed based on player game logs sent to a server that could train an AI based on pouring through logs of actions that human players have taken in certain circumstances. Let a learning model sift through thousands of player games and then create a theory of the game for itself. I bet if an AI did something like that it could play in the top 20% of players, and then the devs would have to work on ways to kneecap the AI on lower difficulties instead of constantly giving it free stuff on higher ones.

Also, and actually decent carrot to convince players to play with their online mode on, and connect their profile to the 2K network, or whatever. Connect your game profile and you will gain access to the good AI player, and help make it even better.
 
I can imagine that some learning model could be constructed based on player game logs sent to a server that could train an AI based on pouring through logs of actions that human players have taken in certain circumstances. Let a learning model sift through thousands of player games and then create a theory of the game for itself. I bet if an AI did something like that it could play in the top 20% of players, and then the devs would have to work on ways to kneecap the AI on lower difficulties instead of constantly giving it free stuff on higher ones.

Also, and actually decent carrot to convince players to play with their online mode on, and connect their profile to the 2K network, or whatever. Connect your game profile and you will gain access to the good AI player, and help make it even better.
I would like to add that machine learning would likely make game development go faster because the developers wouldn't have to spend so much time tweaking coded AI and patching games.
 
PS5 game Gran Turismo got a limited time option to race against an AI. It's different from the usual game bots as it is capeable of reacting to the player and actually plays like a human. Players have reported that it is much more fun to play against that kind of AI. Difficulty levels are solved by giving the player a faster car than the AI with the most difficult setting giving both the AI and human identical cars. It's limited time presumably because it's resource intensive as the AI is not running on the PS5 itself but on a server. (I might be wrong)

So yeah, not only is machine learning possible in video games but it is actually on its way.

Now having such a system in Civ is another beast entirely. Problems I see:
1. Complexity. Having the AI to train on the same track/car while giving it only a handful of mechanics to influence is probably easier than a game like Civ where the map is random and there are a plethora of ways to proceed.
2. Victory conditions. Crossing the finish line first is a clear goal for the AI with many options the AI might favor only one or two.
3. Historical flair. Lets face it, current Civ AI is designed to have some historical flair and “pulls punches” in certain areas just so the player can enjoy all the aspects of the game to its fullest. Sometimes the player doesn’t know what victory condition to go for until half the game. Some people play it just for fun.
4. Steamrolling. Inherent problem of the 4x genre I guess. Players usually start behind the AI and by correct game decisions get ahead by the mid to late game. Having that the other way around seems crazy.

Perhaps Civ 7 could include AI similar to GT so it would be an additional difficulty. Perhaps they could leverage it in another way like only in warfare.
 
I've been amazed by the AI released since a few month, and I've been using both Stable Diffusion (for images generation) and ChatGPT since a few months, but I don't think we'll see one that could efficiently play Civ and be used on a home computer anytime *soon*.

And I don't think that even a server-based AI could do it.

Now such an AI on specific tasks, yes, sure, like role-playing diplomacy maybe, or deciding a global strategy. But what would be the cost to access it for players ?

OTOH, what I can see it be used quite intensively for right now is modding.
 
A generative model is said to be generative because it generates a plausible output given some input. A simple example would be a handwriting generator. This generator would model the probability distribution of pixels (output) given some text (input). When you tell it to write the letter "A", it will sample from the distribution it models to give you a random set of pixels that resemble "A". It doesn't have to walk through a logical step-by-step process of how to construct the letter like a human might (e.g. you have to draw a diagonal left and downwards, and then another diagonal starting from the same spot but to the right and downwards, then join the two diagonals somewhere in the middle with a horizontal line). There are tasks where this sort of approach works fine. Making a generative model play Civ like a human may fall under that, but there's a big difference between playing the game like a person and playing it well.
 
I just tried out chat gpt and it gave me so much complex answers on many subjects that it made me feel obsolete. Seems like there's software that is starting to make machines replace smart people that teach.
 
tldr: The AI isn't good, but the simplicity of the game, allowing the player to optimize their play, is a bigger problem.

The AI isn't so much a problem as is the simplicity of the game. Civ6 made the path through the game trivial. Keeping above the threshold of a city becoming unproductive is really very easy. As long as a city has production, it is fine, and you can easily make more production, and all the required currencies from there with appropriate tile improvements and districts. It isn't a very complex mental process to make decisions on what to build next, and it doesn't change from game to game.

In Civ7 they have to make city building far more complex. They have to reintroduce public health and overcrowding, happiness and loyalty needs to be far more punishing. The calculation to work out what to build next has to be far more difficult or include some RNG so that the computer has a chance on an equal footing. What really upsets people is the AI getting advantages in combat and loyalty to create competition. It isn't that the AI is bad, it is that players can too easily play the game in an optimal fashion. Any game that can be won by a player will be won, and there is little that a different map or Civ provides in complexity.

When I lose a game these days, I pretty much figure I lost because the game was unwinnable, not because I made any particular mistake. I just got a bad start on the map and got overwhelmed by the surrounding Civs with their start advantage. The Civs around you sacrifice themselves, so that an unmolested Civ on the other side of the map can win. That is just no fun.

The game has to be complex enough, so that the player can be lured into mistakes the AI wont make, because it can work out the correct next building to build. There has to be more choice. So for example, the bonus hex should be a pasture, allowing a player to improve it to horses for military, oxen for more production, or cows for more food. Buildings in districts should be useless if you don't have the population to work them and so on.
 
anybody hear about this open source AI called Cicero?

They used it for diplomacy which is notoriously hard for AI to do. I can only imagine how awesome this would make civ if true

 
Yeah, I knew about that.

I think people are getting it twisted. A game like civ doesn't just have 1 AI, it has at least 3:
- A tactical AI for troop movements and tile/improvement management on the map
- A governor AI that assigns citizen tile management, does build orders, etc. (probably 2 separate AIs right there, actually)
- A diplomacy AI that handles trade, war, congress, etc between the other AIs and the human player.
 
Top Bottom