CivVII AI

[Lazy sweeper]

Automation would play a major role. Since 1UPT Ai struggle to move around the troops. It moved archer on front lines instead of infantry. Never got bombardment right. Not even civ V that Ai actually built planes used them right.

Devs should focus on groups movement, like AoE, select a bunch of units, and tell them to perform an action. No anymore a unit by unit cmd inputs.
Human players would get the advantage of moving an army with one click. The group, should be able to move around a bottleneck, all units.
Enmbarkment for all units should be banned, except for Scouts maybe with some kind of perks. So we don't have AI that moves units on water and get crippled as a consequence of
this stupid action by a competent defender. AI would be able to use this group characteristic and pose significant advance in overall tactical aspects of the game.

If they do not start contempleting these changes, there will not be any kind of betterment I'm afraid.

 

[DJ_Tanner]

I would expect the AI to feel better in Civ VII, if nothing else simply adding in the age mechanic would make the AI (even if untouched) stronger. A huge issue with the AI previously was simply the goal was too large/nebulous and it would have the AI making odd choices as 'winning' was an ill-defined and ever-changing prospect with turn 1 being so far removed from turn 500. With ages you can give the AI smaller goals and with the addition of multiple victories within each age the AI can be, if nothing else, far more consistent within the actions it takes.

 

[Malachi256]

I realize it's almost certainly apples and oranges... but it does amuse me a bit that in the era of LLMs and AI really becoming a household tool, that we don't have higher expectations (nor witnessed results) of improved AI in PC games.

 

[AriochIV]

I realize it's almost certainly apples and oranges... but it does amuse me a bit that in the era of LLMs and AI really becoming a household tool, that we don't have higher expectations (nor witnessed results) of improved AI in PC games.

Generative AI and game AI are very different things. For one thing, there is no massive data set of game moves to train on for a game that doesn't yet exist.

 

[Malachi256]

Generative AI and game AI are very different things. For one thing, there is no massive data set of game moves to train on for a game that doesn't yet exist.

If it's just data we're missing, even for an in-dev game, it should be easy enough to use AI vs. AI simulations and/or AI vs. Tester games to get the ball rolling.

It strikes me more as a 'lack of will' than a 'lack of way' issue.

Once the game is out, within a week or two there would be tons of good data they could scrape to improve an AI that had some rough edges.

 

[AriochIV]

If it's just data we're missing, even for an in-dev game, it should be easy enough to use AI vs. AI simulations and/or AI vs. Tester games to get the ball rolling.

It strikes me more as a 'lack of will' than a 'lack of way' issue.

Once the game is out, within a week or two there would be tons of good data they could scrape to improve an AI that had some rough edges.

That's not how generative AI works; it doesn't "learn", it has to be "trained." It's not like a genetic algorithm that can learn by playing itself.

I suppose it's true that if Firaxis made AI their top priority, they could create a Google-class AI that could probably beat most players... but there's nothing "easy" about that. It would probably cost more than the combined budget of the entire rest of the game. As in... a lot more. That's just not a reasonable expectation.

I hope for better AI that does a better job at challenging the player. We'll see.

 

[Lazy sweeper]

That's not how generative AI works; it doesn't "learn", it has to be "trained." It's not like a genetic algorithm that can learn by playing itself.

I suppose it's true that if Firaxis made AI their top priority, they could create a Google-class AI that could probably beat most players... but there's nothing "easy" about that. It would probably cost more than the combined budget of the entire rest of the game. As in... a lot more. That's just not a reasonable expectation.

I hope for better AI that does a better job at challenging the player. We'll see.

We just need a Sinclair spectrum grade capable AI, we don-t need an nvidia driven AI that can take us to Mars and come back alive...

Make your targets straight... We just want an AI that is barely capable of moving its units, have some build priorities, and have a really basic diplomatic understanding.
Ai vs Ai, generative AI, are you serious? There has been nothing like this for ages, and everything worked, and now we should believe its impossible to have a decent AI
bacause it would bankrupt 2K?

I'm out of words for your disillusion.




We have a set of rules for Ranged units, they have a rule that says that prioritize movement in the front lines of infantry units, and ranged should never go in front.
Same for cavalry, except these should priorites lateral movement, and only in some cases, promote brute force attack of enemy lines.
Every unit type draws a field of projected possible attack values, and defence values, and then has to move its units accordingly following this
basic mathematical rule. This very similar to chess Ai.

Ai has a few tricks up its sleeves that can know in advance the defence values of all your cities. that is how it worked in civ 3.
It then decides to go attack the city with the lower defence value, with a big enough force, to overcome the odds.
1Upt made things more difficult, but that was their choice, how they do it, should not be our affairs.
We should be just pointing out the obvious flaws of each design choice, without ever buying into
any narrative that involves generative Ai.

 

[Gedemon]

We just need a Sinclair spectrum grade capable AI, we don-t need an nvidia driven AI that can take us to Mars and come back alive...

Make your targets straight... We just want an AI that is barely capable of moving its units, have some build priorities, and have a really basic diplomatic understanding.
Ai vs Ai, generative AI, are you serious? There has been nothing like this for ages, and everything worked, and now we should believe its impossible to have a decent AI
bacause it would bankrupt 2K?

I'm out of words for your disillusion.

I think the comment was made in the context of generative AI and why it's not the right tool for the problem*. There will be a cost for the player too, using models resulting from such training locally usually requires a very good GPU.

There is no doubt the AI can be improved with normal dev time (and it did during civ6 lifespan, even if it was not enough IMO), but I do think the simple reason they don't put a lot more resources into it, is that the game sells very well with the AI at its current level.

*still pondering if a LLM could be called in some part of the decision process, for role-playing AI, in a mod

 

[Taefin]

At least regarding warfare in VI, capturing walled cities was a very complicated task. However if walls were significantly weakened by mods, then AI often just sat next to cities without attacking them, even if that would lead to capture (perhaps still behaving as if it needed a full siege before attacking). GDRs also just sat next to cities without attacking, which was a shame since they at least balanced human aircraft, but only if there was pressure to stop them. I think player fun dictates being able to outplay a larger enemy, within reason, and my biggest hope for AI is that the player can be pressured by AI conquest (of other AI) and feel an urgency to intervene even when outgunned. VII looking simpler makes me hopeful (eg auto unit upgrade on era) that even just a second person on the AI team to bounce ideas off of will be enough.

 

[Arcaian]

If it's just data we're missing, even for an in-dev game, it should be easy enough to use AI vs. AI simulations and/or AI vs. Tester games to get the ball rolling.

It strikes me more as a 'lack of will' than a 'lack of way' issue.

Once the game is out, within a week or two there would be tons of good data they could scrape to improve an AI that had some rough edges.

Using exclusively ML-based 'AI' algorithms for the actions of enemy units is an incredibly complex thing to do. Look at the amount of possible move options your starting warrior can take before turn 20 - I'm pretty sure it's into the billions (as you have to include all the redundant moves as well - the algorithm has no knowledge of why it makes no sense to move one tile north-east and then back south-west). If you're advocating using machine learning to figure this out, it will need to have data for all of these moves, and it will need some metric by which to judge their success. On the first point, it it tremendously expensive to get this data - you can't use AI vs AI simulations for the vast majority of this, as then all the people in the game will be doing useless moves where they just move their warriors back and forth for no reason. On the second point, what metric defines success here? Do you go for the big one, and say it's whether the AI won the game? There are so many factors that aren't related to the warrior's movement there that you'll get a huge amount of confounding factors - you'll never optimise your warrior's movement that way, not with any plausible amount of data. Do you go for something like the number of tiles revealed? You do use your starting warrior to explore much of the time, but then it'll encourage the AI to just go off in one direction into the fog of war without a thought, leaving their cities exposed. Do you go for something like barbarians defeated? Then the AI won't explore with it. Do you generate one AI for the warrior based off of barbarians defeated, and one for the scout based off of tiles explored? Now you need to double the cost of all of this work.

And on top of all of that, you end up with a fundamentally unmoddable AI. If all of this is trained off of data that has Scouts have 2 movement points and you give them 3, that might break the AI. It might not - it's almost impossible to know what and how the AI has correlated, and what changes will break it. Can you do a mod that creates a new type of terrain that imposes a combat malus - maybe reefs for water tiles if they don't still give a bonus? Well you could, but the AI probably won't consider it when making a decision. Maybe the AI was directly exposed to the combat penalties/bonuses of a tile when training, rather than just learning that hill = more defence? In which case you might not be able to have the AI react to a modded unit that has an aura lowers the combat strength in certain situations, or one that heals units, or anything like that - you might be able to do it, but it'd be a giant pain to find out. It's also a difficult-to-patch - any time you change anything on which it was trained, you'd need to re-train it until it behaves appropriately, so it adds significant complexity and cost to any patching efforts. Using ML-based algorithms to train a model for Civ's AI doesn't create something that is actually capable of thought or understanding - it just generates a bunch of correlations between the data on which it was trained. It's not plausible to use it for anything in Civ except, perhaps, for very specific situations with a very clearly defined short-term goal, and efforts would be far better placed in actually just hiring a few more game AI people and improving the logic used - and in designing the game to work better for the AI, as they have done.

 

[GoneFishingX]

The idea that a Nintendo Switch or even a minimum spec PC could do all the parallel processes necessary to handle a super complex computer player is unlikely and pointless.

This idea that somehow computer opponents should ever be daunting due to their skills alone is stupid. It's just a total waste of dev time. The CPU opponents in every civ game are dumb as rocks and only ever posed a challenge because of massive bonuses and opaque systems.

The thing with mechanics like 1UPT, external districts, and worker charges is that it makes the computer's bad decisions more obvious and more costly. When they place a district in a really suboptimal spot, it's obvious. When they manage their 1UPT armies poorly, you can easily see many smarter moves because you likely prepared for them.

In CIV 4 (where I have most of my experience) wars are basic and boring. Like 99% of the map is unused as waging war involves moving a doom stack to the closest city and camping there. It's really easy to make a computer manage that and they still regularly fail to pose a meaningful threat. Those stacks also allow the AI to easily bring their production bonuses to bare.

TLDR - The AI has never been smart and there is no real reason to make it smart. The systems were just masking their poor decisions better by having less points of failure.

 

[Lazy sweeper]

The idea that a Nintendo Switch or even a minimum spec PC could do all the parallel processes necessary to handle a super complex computer player is unlikely and pointless.

This idea that somehow computer opponents should ever be daunting due to their skills alone is stupid. It's just a total waste of dev time. The CPU opponents in every civ game are dumb as rocks and only ever posed a challenge because of massive bonuses and opaque systems.

The thing with mechanics like 1UPT, external districts, and worker charges is that it makes the computer's bad decisions more obvious and more costly. When they place a district in a really suboptimal spot, it's obvious. When they manage their 1UPT armies poorly, you can easily see many smarter moves because you likely prepared for them.

In CIV 4 (where I have most of my experience) wars are basic and boring. Like 99% of the map is unused as waging war involves moving a doom stack to the closest city and camping there. It's really easy to make a computer manage that and they still regularly fail to pose a meaningful threat. Those stacks also allow the AI to easily bring their production bonuses to bare.

TLDR - The AI has never been smart and there is no real reason to make it smart. The systems were just masking their poor decisions better by having less points of failure.

Well, not entirely pointless.
Maps are evolving, vertically, so units should follow the least amount of resistance to reach the nearest city, not just a straight line.
So it is likely ok for devs to experiment with various geometries for all kinds of units and possible overlapping/combination effects,
that could bring significantly better results over time.
Chess evolved over who knows how many thousands of years before settling with a defined set of rules, yet civ has already
changed its basic rules three or four times from its inception. And on each new iteration it's almost like ground zero.

I repeated this multiple times. Get your rules straight!
It's pointless to evolve a working Ai with this set of rules, if then Civ 8 will change all of them...
This is the third episode of 1Upt, and there is no general consensus of what rules should all units obey...
let alone have a capable Ai...

want an example? roads and Ai is difficult to program, thus, roads are removed.
parallel processing idk to what amt, but some basics should be covered.
the grunt of the work could be done from bigger gpus and Ai, and then optimize the best set of rules
to work on all but the slowest of Switches out there... its work.. it is not easy... but it is not pointless either...

 

[The_goggles_do_nothing]

If there was a client api for civ someone could write their own AI.

 

[nudden]

Can they pls just hire the guy who wrote the code for the Civ 5 Vox Populi AI?
It's not perfect but at least this AI can play the game and give you a fight ...

 

[stealth_nsk]

Can they pls just hire the guy who wrote the code for the Civ 5 Vox Populi AI?
It's not perfect but at least this AI can play the game and give you a fight ...

There's a myth about allmighty mod developers who create purely fantastic code and those developers working for corporations, who can't write code at all.

In real life things are a bit different:

1. Mod developers don't count their time and are not restricted by any deadlines. They often put way more efforts into modding than was put in the development of the appropriate part of the game. If hired, they have to follow company deadlines as those who work for the company.
2. Game developers have to ensure their code works on variety of configurations following system requirements. Mods often raise system requirements to skies and crash in case something goes wrong.
3. The quality of game AI is not a single good-bad slider, it has to give challenge appropriate to the selected difficulty level, while at the same time provide some role playing. Not everyone actually like Vox Populi AI.

So yep, the experience of hiring mod developers results in pretty expected outcome - they don't perform better than experienced game developers, usually worse. Just a reminder - vanilla Civ5 was led by one of the most respected Civ3 modders and its vanilla AI was written by him.

 

[GoneFishingX]

Maps are evolving, vertically, so units should follow the least amount of resistance to reach the nearest city, not just a straight line.
So it is likely ok for devs to experiment with various geometries for all kinds of units and possible overlapping/combination effects,
that could bring significantly better results over time.

I'm not sure what you mean here. A* and other pathing algorithms have been around since the dawn of games. We know how to make a computer calculate the fastest path to something. The problem 1UPT makes for pathing in civ is how easy you can disrupt it, plus moving 20 units requires 20 actions instead of 1 like in civ 4.

The core issue with Civ is that once 1UPT showed up the computer player had to exponentially increase the number of actions it had to do in a turn while still not taking too long.

When it comes to keeping consistant rules that doesn't really fix anything. If we just look at army behavior from Civ 5 -> Civ 6 I don't see much difference. Sure districts add complexity to attacking, but the base rules of movement and how to successfully attack is the same. Computer opponents in other games pure combat games (Advance Wars, Unity of Command, Panzer General, etc.) handle 1UPT grids with ease.

The reason your computer opponents seem to fail more often to successfully attack cities (personally I don't think it's changed between Civ 4 -> 6) is because;
1 - having them actually succeed isn't necessary to make most player feel tension and if they actually succeeded the player might get frustrated and quit.
2 - there are more points of failure causing suboptimal decisions (if the AI fails 1% of the time it will happen more when there are more actions)
3 - to take less time and create a smooth experience on all hardware (civ is awesome for being able to run on really old machines) AI has to 'forget' about more units as it can't stack them.
4 - because Civ has decided that all actions take place in the same layer (scouting, improving, and combat) it means 1 unit could be doing multiple things and get into friction between those roles or even bump into units doing those roles and be locked out of actions.

I think Humankind had the right idea with how it managed war because it allowed for the separation of combat AI and over-world AI. The commander system in CIV 7 should help with making the AI have an easier time moving an army.

 

[Lazy sweeper]

If there was a client api for civ someone could write their own AI.

I am being a bit out of touch lately, beg my perdon.
I am writing an essay on racism and multicultural society and I am breaking down.
It can be frustrating to be an Ai dev and have to go through all this nonsense I
have spammed around. This is just about me, I am not referring to other people on this
thread here. Or others.