Likelihood we'll get a decent AI

[DeckerdJames]

Yes. It does. And quite a bit more than that.

A single 4-speed horseman unit standing on flat grassland in all directions can move to 61 tiles (this includes not moving at all). Six of these horsemen can move in 61^6 = 51 billion configurations. So we're not looking at a standard map, but rather just three cities worth of land (assuming there's some overlap between cities, otherwise it's just two).

Meanwhile, a chess engine that looks forward 8 ply has 20^8 = 25 billion possible configurations. Note that 20 is an approximation of the average number of legal moves in a given chess position - the precise number may vary.

And that's just a game aspect that can easily be quantified in terms of possibilities! How about 'when to declare war', 'what of these 8 units, 8 buildings, 6 districts and 6 wonders to build next' (28 options per city, at 8 cities this by itself beats 8 ply as well), 'where to settle next', and so on?

The only way to build a 4X AI is by generalizing things like this so that you never even consider the options you don't take. Simply considering and discarding bad options like chess engines do just doesn't work, it's too slow.

Civ isn't chess though. These are fundamentally different games. Chess is a game where every piece can capture a whole piece in 1 turn. Civilization is a game where pieces do HP damage. The game doesn't need to know how many ways or in how many different move orders the 6 horsemen can reach a particular hex. It just needs to know which pieces can be attacked and how much damage they can do.

You could even go so far as to say that by simply evaluating the maximum amount of damage the 6 horsemen can do within their range is a good place to start. Now the damage might vary depending on the strength levels of the opposing pieces. For example, a horseman attacking a warrior might do much damage and take little in return, but it might also mean that the warrior poses less threat to the horseman during the coming counterattacks.

So, without sitting down to design the system here and now, in general I would say that a threat assessment should be done to determine which pieces poses the greatest threat during the coming counterattacks, and what are the optimal attacks to make to reduce the damage that will be sustained from the counterattacks. Now that is only one step. Of course, the horsemen can attack their optimal targets from multiple hexes if they have the movement points. Another assessment could be made to determine which hexes are the optimal hexes to occupy. In civ, the defense modifiers are a simple mathematical factor, but if the AI was ending their turns in the best defensive positions, the AI would be doing ok. However, they might also want to slow down the advancement of enemy units. Strategic concerns like this are evaluations made on the geography and infrastructure, like roads, but also water which just a medium for travel. The point is, that good spots on the map can be identified and the computer can play positionally.

In a game like Civilization, with such a large board, if it plays positionally it will already be playing a much smarter game even if didn't look 1 ply ahead. It is the same in chess. Once upon a time in chess history, positional chess was new.

So, what is important? Defensive modifiers, movement speed of terrain, therefore roads and water, water movement changes based on technological advancements, resources, districts, improvements that can be pillaged, proximity to enemy units can affect the evaluation of a tile, proximity to a strategic goal such as a city center district, and the list goes on. You could evaluate the board for good positions and already the AI is much better to take those positions and to move across the board from good position to good position rather than as quick as possible. (unless strategic concerns make speed more important, but that results from evaluations of a higher order than brute force searching)

Also, the close formation of units with the military tradition civic allows the defensive bonuses to increase, but these can also be added to the bonuses from the geography. For example, the computer can easily evaluate that a patch of hilly terrain is better defensively than flat grassland but slower to move across without roads. However, when multiple units are grouped together, the defense rating goes even higher, so even a pack of warriors in wooded hills can pose quite an obstacle to a stronger man-at-arms unit. So, the AI can be programmed to evaluate positionally for a group of units rather than just for each unit individually.

Suppose that the AI wants to siege a city with siege units. It can choose the best positions for those siege units long before they arrive and group them with support units. Even if they were warriors surrounding it, they would lend +12 to its defense and if there was a good, wooded hill, it would +18 over normal against the city's defensive bombardment. Of course, adding six support units to the siege unit is an ideal number for the siege unit's defense. A good AI has to be able to identify the important units. This is simpler than chess. A pawn is generally worth less than a queen, but a pawn can deliver checkmate. In Civ a warrior is worth far less than a tank. The strength numbers make it easier. Among units that are relatively close in strength there is some consideration to be had based on the overall objective.

The point is a game like Civ is far more ideal for positional play that brute strength. Brute strength is still going to be a part of it, but it doesn't need to run through 50 billion moves.

thoughts?

 

[DeckerdJames]

It's worth noting that even if they did brute force it 8-10 turns deep, that would be woefully insufficient for a game like Civ where strategies routinely have to be in mind 100+ turns in advance; brute forcing is simply untenable for a game like Civ in any meaningful way (even a duel map forced into only one city for both AI would be functionally impossible). As you say, it also isn't necessary; the current general model of saying "this civ is better at science from its abilities, so it'll have 1.5x weighting for science victories/infrastructure/etc. It then spawned in a great spot for campuses, so it'll take advantage of that" is a pretty sane way to do it. There's plenty of work to be done in making those weights more effective for substantially improved AI behaviour (Real Strategy for Civ 6 shows that), and just in abstracting strategy into these weights in a way that works more effectively too. The down side of this system is that it can be quite hard to fiddle with the weights - changing the AI's weighting for campuses doesn't just give them a bit more of a focus on beakers (as we saw with that accidental reweighting of campus building priority in civ 6), it started bankrupting the AI, they built no units, and the units were stuck several ages behind their actual tech, just as an example. Putting in a "player customisation screen" where they can just re-weight these themselves is going to be pretty tricky, though having a few options ("more aggressive AI", "more untrustworthy AI", "more player-like AI" (i.e. pursues victory over roleplaying)) would certainly be do-able as they could be pre-weighted.

This whole discussion gets at the problem with "just make the AI smarter on harder difficulties, not get unfair advantages" - it's extremely difficult to do. If you make the AI more aggressive on Deity difficulty, you can't just bump up some AI_WEIGHT_MILITARY_AGGRESSION value, as there are a huge amount of flow-on effects. You could tweak and tune the parameters for each difficulty so they work in concert with each other, but that's a lot of work - and then on top of that, you'll have to reconsider those weightings pretty regularly during development, patching, and expansions, so the work is on ongoing maintenance concern. The most efficient way to do this would probably be to have a mode where no rendering occurs and you can just have AI automatically play against each other and used directed evolution techniques to alter the weightings based off of the AI that win. That could be substantially automated ... and then runs into the issue of the AI being hyper-optimized for beating another AI, and bad at reacting to a player interfering in their normal plan. This is all much harder to do than one would think, and IMO the best solution would be designing the game while keeping the knowledge that the AI are going to have to use these mechanics as well in mind; some of the mechanics are unnecessarily difficult for the AI to navigate.

What Leyrann and I were getting at, I think was how the computer handles units in combat and on the board in general, but your approach to AI brings an important point up. There is more than one AI at work. I am pretty sure that is the case in Civ 6, but your entire discussion here is a different sort of AI than the one that would handle unit combat. There are many systems that have to work and together they hopefully look like an intelligent human.

 

[GeneralZIft]

This is quite possibly the most divisive topic about AI that you can find on these forums. Some people argue as you do, some people argue that it's the worst thing ever for the AI to consider anything victory-related in their decision-making.



I'm pretty sure people complained about the AI doing this in Civ V quite a lot, although it's not something that'd reach my personal top 50 biggest issues with the game.



Oh yeah that's why everyone has constantly been declaring war on the USA, Germany, etc over the past 70 years.

1. I am not talking necessarily about Victory Conditions. I'm just saying if a player is fairly unarmed with huge territory and a big lead, then maybe the opponents should be jealous or angry

2. I can't believe this was ever a complaint? This was a good feature to me.
This is what diplomacy is all about. If leaders wrote on their forehead "I love you unequivocally" as they do in Civ6, that would be both unrealistic and predictably BORING.
Maybe the AI could backstab less than Civ5 - but the fear of a potential backstab needs to be there... Otherwise there's no back and forth.

You should have to keep your Allies happy not just coast by on a random friendship declaration. In my opinion.


3. That's besides the point. Not that this is the sole reason, but obviously, in real life, America has a military spending that dwarfs the entire world combined.

Let me summarise the ideal AI in terms of war declaration:

Depends on territory dispute
Depends on personality of AI
Depends on wealth of AI Civ compared to wealth of Player Civ
Depends on Strength of AI Civ compared to Strength of Player Civ
Regards potential Allies in Player's potential enemies where possible
Declares war if the likelihood of winning it's favour, and recruits Allies if need be

(With respect to backstabs):
Weighs the potential benefits of annexing your land VS. the benefits of keeping you around (again with respect to its personality)

I have to say, this is not super complicated stuff - this has been touched upon very lightly between Civ 5 and Civ 6.
They could go a long way just by making these opponents more interactive.

 

[Arcaian]

1. I am not talking necessarily about Victory Conditions. I'm just saying if a player is fairly unarmed with huge territory and a big lead, then maybe the opponents should be jealous or angry

That's actually still a pretty contentious point - it makes sense if you're looking at it as a game, because if someone has a big lead then the best odds of success for everyone else is to work together to take them down. But is that something that makes sense in the narrative that civ is trying to tell? It certainly can be, but not universally - you don't often see smaller states under the US' protection like Australia getting jealous at the US, even when there wasn't another significant global power threatening them. It seems that some of the time, states react with jealousy, and other times they're happy to try and get what benefits they can from working with the global power. If you're focusing on the AI as trying to improve the game side of things, it makes sense for everyone to ally against the leader; if you're focusing on the narrative then some civs should be willing to ally with the global power, and others won't. The issue is that the second almost makes things worse for snowballing, and there's a pretty fundamental disagreement between the AI-Civs-are-playing-a-game perspective and the AI-Civs-are-following-a-narrative perspective.

2. I can't believe this was ever a complaint? This was a good feature to me.
This is what diplomacy is all about. If leaders wrote on their forehead "I love you unequivocally" as they do in Civ6, that would be both unrealistic and predictably BORING.
Maybe the AI could backstab less than Civ5 - but the fear of a potential backstab needs to be there... Otherwise there's no back and forth.

You should have to keep your Allies happy not just coast by on a random friendship declaration. In my opinion.

Again, it depends on one's perspective. If an AI leader is well-known for being deceptive, I think everyone is OK with it - if Catherine De Medici says she has a friendly attitude and wants to engage with you, but is secretly sending spies to mess up your empire and then declares war in the turmoil, that seems pretty consistent with the narratives being told. If Gilgamesh does the same thing, it's going to come off as very game-y - he's meant to be committed to his friends by the narrative the game is telling. If every AI declares friendship and then calculates that the expected gain from betraying that friendship exceeds staying in the friendship and does so, it doesn't feel like it's respecting the narrative being told by the game. I think this one is also an issue with presentation of the system - the information the game provides is almost always authoritative and correct; the military strength/culture per turn/science per turn/etc summaries of each leader are never incorrect, for example. So when the game says "this person is friendly", I think it feels like a more frustrating betrayal for a lot of people because there's a (pretty reasonable) default assumption that all information the game gives you is true.

 

[Arcaian]

What Leyrann and I were getting at, I think was how the computer handles units in combat and on the board in general, but your approach to AI brings an important point up. There is more than one AI at work. I am pretty sure that is the case in Civ 6, but your entire discussion here is a different sort of AI than the one that would handle unit combat. There are many systems that have to work and together they hopefully look like an intelligent human.

I do agree that they're pretty different AI systems, and absolutely could be handled differently. At the same time, I think the core of my point applies to both of them - it's hard to brute-force calculate the next 10 turns of tactical AI to make a decision for the same reasons it is hard to brute-force calculate the next 10 turns of the strategic AI that I focused more on. The system you've described of positional play is in line with my thinking, I think - you use weighting to push for behaviour we want the AI to perform based on our knowledge of the reasonable tactics in the game, rather than try to calculate predictions of the consequences of the action into the future. It's tricky for tactical AI, however, as you get a pretty huge advantage from being able to think a turn or so ahead in the tactical combat minigame. If they just target the hex they can reach/see based off of the hex's value (factoring in your threat assessment calculation), it's really easy to take advantage of that by baiting them into an overextended position. At the same time, if one is trying to focus on positional calculations and not going into the future turn calculations, it's hard to include information like "5 turns ago we saw a bunch of archers nearby, they're probably close to the city we're attacking".

 

[GeneralZIft]

That's actually still a pretty contentious point - it makes sense if you're looking at it as a game, because if someone has a big lead then the best odds of success for everyone else is to work together to take them down. But is that something that makes sense in the narrative that civ is trying to tell? It certainly can be, but not universally - you don't often see smaller states under the US' protection like Australia getting jealous at the US, even when there wasn't another significant global power threatening them. It seems that some of the time, states react with jealousy, and other times they're happy to try and get what benefits they can from working with the global power. If you're focusing on the AI as trying to improve the game side of things, it makes sense for everyone to ally against the leader; if you're focusing on the narrative then some civs should be willing to ally with the global power, and others won't. The issue is that the second almost makes things worse for snowballing, and there's a pretty fundamental disagreement between the AI-Civs-are-playing-a-game perspective and the AI-Civs-are-following-a-narrative perspective.


Again, it depends on one's perspective. If an AI leader is well-known for being deceptive, I think everyone is OK with it - if Catherine De Medici says she has a friendly attitude and wants to engage with you, but is secretly sending spies to mess up your empire and then declares war in the turmoil, that seems pretty consistent with the narratives being told. If Gilgamesh does the same thing, it's going to come off as very game-y - he's meant to be committed to his friends by the narrative the game is telling. If every AI declares friendship and then calculates that the expected gain from betraying that friendship exceeds staying in the friendship and does so, it doesn't feel like it's respecting the narrative being told by the game. I think this one is also an issue with presentation of the system - the information the game provides is almost always authoritative and correct; the military strength/culture per turn/science per turn/etc summaries of each leader are never incorrect, for example. So when the game says "this person is friendly", I think it feels like a more frustrating betrayal for a lot of people because there's a (pretty reasonable) default assumption that all information the game gives you is true.

You need to consider it as all players are "big" players on the world stage. They're all global powers by assumption - it's not that they're all ganging up on the star player, it's that they want to BE the star player.
That's what they should aim for no matter what their personality is like.

AI players that aren't making huge moves... Well... Those are City States.
States like Real life Australia in your example are the City States in-game (so to speak)
The Global Powers are the Civilisations in-game.

I agree with you about information. However I do think that the narrative is better served by you know, actual plot twists instead of predictable personalities.

Again, backstabbing, it should depend on the personality too - so Gilgamesh wouldn't be the type to do it unless the odds were really against him.

I would like to see AI Personalities befriend their Aggressors and Backstab them when they're not looking.
Like holding a grudge. This is one form of backstab I wouldn't mind from a personality like Gilgamesh.

(More like Plotted Revenge than Backstabbing)

 

[DeckerdJames]

I do agree that they're pretty different AI systems, and absolutely could be handled differently. At the same time, I think the core of my point applies to both of them - it's hard to brute-force calculate the next 10 turns of tactical AI to make a decision for the same reasons it is hard to brute-force calculate the next 10 turns of the strategic AI that I focused more on. The system you've described of positional play is in line with my thinking, I think - you use weighting to push for behaviour we want the AI to perform based on our knowledge of the reasonable tactics in the game, rather than try to calculate predictions of the consequences of the action into the future. It's tricky for tactical AI, however, as you get a pretty huge advantage from being able to think a turn or so ahead in the tactical combat minigame. If they just target the hex they can reach/see based off of the hex's value (factoring in your threat assessment calculation), it's really easy to take advantage of that by baiting them into an overextended position. At the same time, if one is trying to focus on positional calculations and not going into the future turn calculations, it's hard to include information like "5 turns ago we saw a bunch of archers nearby, they're probably close to the city we're attacking".

In Civ 6, pretty much everything boils down to a number. For example, adjacency bonuses are a weighting system that helps the computer determine where to build a district. The loyalty pressure from foreign cities are numerical and influence where a city center will be built. I think Civ 6 is a system of weights like you are describing and more could be done with it. For example, the positional play with units could be done as a weighted system like you said. The diplomacy system is based on numbers.

Basically it is a system of scales. Lower to higher numbers. For the units, you probably just create a layer or layers for each turn and add collection of numbers to the tile. Two simple numbers are defensive value and another could represent the strategic value of occupying it. There could many such numbers, so the AI part is in the evaluation and scoring of the tiles. The units just move to the best scores based on what is relevant for them.

On the units there might also be scoring. Strength is a score. Hp. Also you can just have simple switches. You could have a switch for defensive vs offensive. Some other system could set those switches, the unit just moves toward the appropriate scores based on the switch. There could many such switches.

That is fairly basic foundation for positional play. I think there is probably many way to implement a system like that with some imagination.

Brute force comes into play when there will be multiple strikes and counterattacks but that to could also result in scoring the tiles and just be one evaluation that helps determine unit moves. You would need to sit down and work the system out.

I won’t pretend I can come up with a working system in my head, but if you just added up the total damage that the enemy could apply to each tile you have a set of scores that can be ranked and you could determine which tiles a particular unit could definitely die on. If the units just moved to the least dangerous tiles you could say they were practicing a strategy of survival, which honestly could be put on the unit as a switch.

There are cases where move order is important, like removing a zone of control so another unit can move through it on the same turn or deciding to attack with ranged or melee first when the melee unit will occupy the tile if it has the final strike.

I think setting up units with a collection of switches that direct unit behavior produces a system where other systems could set those switches to produce a great variety of behaviors. The benefit is the low level scoring and switching systems can be implemented piecemeal and enhanced over time. It is the systems that set the switches that need to know why to set them.

It allows for something like a morale system, if appropriate to the game, where the unit can set, say the aforementioned survival switch and override other switches and so battlefield panic could set in.

I imagine there are other approaches. This one allows the unit to read the scores and make its own navigation while other systems do the scoring. Another system might be more top down and determine all the moves exactly and deliver a movement order to the unit. Another might allow units to report newly discovered tiles as if it were radioing in intelligence reports which might allow for a different method for top level systems to gather intelligence in the era of radio communications while in the ancient era units must return. I don’t know if that would make sense for Civ. Just thinking about the systems.

Scoring also makes intelligent scouting pretty easy. Scoring the FOW higher, what is revealed low, what is in the mid fog would be low for geography scouting. If anything good is revealed like fresh water or natural wonders, it can rescore the unrevealed areas around them even higher.

Patrols are possible by scoring the mid fog around the borders highest and allowing the unit to report freshly revealed tiles so that the score is lowered for them temporarily, or allowing the unit to adjust the score directly.

Leaders and civilizations could score differently. One leader might love natural wonders so it scores them highly for scouting and settling and defending. I think Civ 6 already does some of that.

 

[ThunderLizard2]

This is irrelevant. Aside from that, you and me as consumers and players do not want a good AI. Studies show we automatically think a good AI is unfair and/or cheating.

Speak for yourself. What I want is an AI that can play the game - not just get cheat bonuses.

 

[ThunderLizard2]

I think that the goal for Civ 7 is an AI that can challenge and win against deity level players without needing bonuses.

Agreed. But I'd take an AI that has a few bonus like visibility advantages. The AI in Civ VI was just garbage and they failed to release dll code so modders could fix it.

 

[Vincent_lacruz]

Speak for yourself. What I want is an AI that can play the game - not just get cheat bonuses.

More important would be to have an AI that can be relevant for all difficulties levels, without bonus.

That could be done with weightening ponderations: computing the possible strategies would give points between 0 and 1. For deity level, the AI would chose the result in the range 0.95-1, noob level from 0.2-0.4, etc... That way a low level AI could sometime do good plays, and deity level somtimes below average plays.

But this is easy to say, and far more difficult to program.. Because you must not only have an AI that is able to move a few units, but one that should be able to mix between short term objectives and long term investments... Building units for the sake of building units is useless if that prevent you from developping your infrastructure. Samewise being able to hold defensively just barely against an agressive opponent the time to develop yout technologies to get that win (or to outclass him) can be decicive... Those kind of comportments are hard to code...

 

[Leyrann]

More important would be to have an AI that can be relevant for all difficulties levels, without bonus.

I feel like a broken record here.

This is not feasible. Games in the 4X genre are too complex for an AI to be able to match a Deity-level player without bonuses. End of story.

 

[GeneralZIft]

I feel like a broken record here.

This is not feasible. Games in the 4X genre are too complex for an AI to be able to match a Deity-level player without bonuses. End of story.

People used to say this about RTS, and AI for that genre became much better. Strategy games have a long way to go, but one easy way would be to code in build orders and some more intelligent algorithms for combat and expansion

And your AI could cheat but you should make it less noticeable. For one, they shouldn't start with 100 settlers and warriors; that is too blatant.

Maybe also, they shouldn't have higher science and culture yields. If they do, don't tell the player.
Maybe instead, their production and growth is easier, but again, this way it's not too noticeable.

The point is not that the AI cheats the problem is that the AI is so blatantly and obviously cheating right in front of you... And in the end it doesn't even make them good enough to win.

A combination of better algorithms, slightly more aggressive and opportunistic leaders, and very subtle cheating, will make for a better experience overall.

 

[Leyrann]

People used to say this about RTS, and AI for that genre became much better.

I've recently been playing Red Alert 2. RTS, release date 2000.

It might be because I'm less experienced at it, but that AI genuinely feels harder to play against than Civ VI Deity for me.

They didn't become better. They've always been better. Because unlike in TBS, in RTS speed matters.

In a TBS game, I don't miss units attempting to enter my lands, because I get to see them walk closer during AI turns, or failing that I'm checking my borders in my own turns. In an RTS game, I'm focused on one area, and an enemy unit enters through another place.
In a TBS game, I can deal with several frontlines at once, because I have to give orders once a turn. In an RTS game, I keep losing units (or structures, on defense) because I'm busy with another area of combat.
In a TBS game, I can individually control who each of my units is attacking. In an RTS game, I have to give general orders.

These, and other aspects, all provide extra difficulty for the player, but not for the AI - an AI can do all these things in an RTS just fine.

Obviously, the goal should always be to make the best AI possible, I fully agree on that. I also agree that in Civ VI, some bonuses (in particular the game start ones) are too blatant. But I want the end result to be that a Deity AI can at the very least give me a serious challenge - and perhaps Deity should even be too hard for me to beat, as I am by no means one of the best Civ players out there. You get as close to that as you can manage with coding, and then have to somehow cover the rest of the distance with bonuses.

 

[GeneralZIft]

I've recently been playing Red Alert 2. RTS, release date 2000.

It might be because I'm less experienced at it, but that AI genuinely feels harder to play against than Civ VI Deity for me.

They didn't become better. They've always been better. Because unlike in TBS, in RTS speed matters.

In a TBS game, I don't miss units attempting to enter my lands, because I get to see them walk closer during AI turns, or failing that I'm checking my borders in my own turns. In an RTS game, I'm focused on one area, and an enemy unit enters through another place.
In a TBS game, I can deal with several frontlines at once, because I have to give orders once a turn. In an RTS game, I keep losing units (or structures, on defense) because I'm busy with another area of combat.
In a TBS game, I can individually control who each of my units is attacking. In an RTS game, I have to give general orders.

These, and other aspects, all provide extra difficulty for the player, but not for the AI - an AI can do all these things in an RTS just fine.

Obviously, the goal should always be to make the best AI possible, I fully agree on that. I also agree that in Civ VI, some bonuses (in particular the game start ones) are too blatant. But I want the end result to be that a Deity AI can at the very least give me a serious challenge - and perhaps Deity should even be too hard for me to beat, as I am by no means one of the best Civ players out there. You get as close to that as you can manage with coding, and then have to somehow cover the rest of the distance with bonuses.

Actually, the AI did get better over the years. Glad you mentioned Red Alert 2 actually - that AI does cheat. It sees the entire map at once. It gets additional income at the start. It can produce units without having the required prerequisites.

Speaking from personal experience (I am an RTS modder), I've seen the improvements of AI in the RTS genre.

So you know, and this is somewhat applicable to 4X, in this case, the old RA2 AI plays 'braindead' and just spams units with little strategy or thought.
RA3 AI for example, changes its playstyle depending on income, total money, total army, etc.

So for 4X, especially Civilisation, there is definitely some space for improvement. It can't be just me that feels that the AI is on some kind of autopilot, floating off of the cheats it uses and otherwise is kind of sleepy and boring.

Basically, what I'm trying to say is that Civ6 is way behind in terms of AI compared to other games and other genres, and it's meaningless to just blame it on the fact that it's a strategy game.

 

[Voidwalkin]

Actually, the AI did get better over the years. Glad you mentioned Red Alert 2 actually - that AI does cheat. It sees the entire map at once.

The difference between the two genres is all about the time the player gets to make decisions vs AI. Player has unlimited time in 4x. The AI actually doesn't; all its decisions must be made quickly enough that these two things are true
A: cpu doesn't crash
B: the player is not bored to death by hitting end turn, because the AI is running millions of calculations. Paradox games have this problem. At a certain point, AI decisions and calculations become so overloading that late game lag becomes a core issue of gameplay.

AI has to play bullet chess, while the player gets unlimited time - this is an inherently unfair advantage that is impractical to change.

 

[GeneralZIft]

The difference between the two genres is all about the time the player gets to make decisions vs AI. Player has unlimited time in 4x. The AI actually doesn't; all its decisions must be made quickly enough that these two things are true
A: cpu doesn't crash
B: the player is not bored to death by hitting end turn, because the AI is running millions of calculations. Paradox games have this problem. At a certain point, AI decisions and calculations become so overloading that late game lag becomes a core issue of gameplay.

AI has to play bullet chess, while the player gets unlimited time - this is an inherently unfair advantage that is impractical to change.

Practically speaking, the player does not spend forever making decisions, and the AI has the capacity to make more informed decisions than the player could ever make.

Only the AI could technically sort all the tiles in its entire empire by potential yields and then decide to improve the highest potential for example. That is, it will always make the best choice, and it takes 20ms to make that choice, for example.

Anyway, the AI doesn't need to play as good as the player. It just needs to be convincing and entertaining.

 

[Lazy sweeper]

How are you going to decide which one of 6 units, 8 buildings, 6 wonders and several projects and districts you're going to build in City #1? And what about City #2? City #3? And building something in one city affects building it in another - if I'm already building an Archer in City #1 I might be better off building a Library in City #3. If you've got ten cities, that's 20^10 = an enormous number of permutations

I'd try to isolate them in multiple Ai. One solely for movement, strategy. Another for all other things related to micromanagement. Just like I would like to have a General, (or more...) that could pursue one tactic (i.e. Conquest or Domination victory) but the 'other' Ai, (Governor, Industrialist, Civic engineers, etc) are pursuing other victory types (Cultural, Science, etc). Then could start a civil war, where the Army would split in loyalists to the 'Crown' or 'Republic' and the 'Rebel Army'. The outcome of the civil war could trigger an Authoritarian leadership, or on the contrary a 'God like King' Empire.

Then try different iterations of multiple Ai within a single civ, vs one main Ai (maybe Governor, maybe General) still limited in its faculties, whilst the other aspects of the priorities list would follow the traditional coded way...
Given the outburst of Ai capable gpus, e-gpus, and the like, inside almost every single modern chipset, external Ai powered gpus, etc, someone has to mine this power, today chess Ai on any Apple device is so unbelievingly powerful that it could shred Blue to pieces. If only One Civilization VII Ai could do one tenth the calculations that Apple Chess Ai is doing right now, it would be already quite extraordinary.

I mean, is it there any usage of Ai currently in Civilization or in AoE that could arguably be pre deterministically compared to Chess Ai?

 

[Lazy sweeper]

This looks fun.... Chat GPT Civ III Ai??
I still have to watch this... don't know what to think...

OK...mmmmh... interesting... Chat GPT prioritise trading tech???
I guess it's something important then

 

[UWHabs]

The first goal of any computer AI is making sure it doesn't make dumb moves. Like the early civ 6 days where it could send enough units at you to capture a city, and then would swap around units on tiles instead of shooting the city, or would start floating a catapult in the water because... I don't know why. You want to make sure if its city is under siege, it's not going to keep building a settler.

The hard part is that for any AI in a game as complex as civ, there's just so many pieces to juggle, for both short term and long term strategy. It has to know what its long-term victory goal is, what it might do if it struggles for that, and for each victory goal, figure out how to get there. Like as a human, I can see sometimes that I might be aiming for a science victory, but suddenly I will go build 5 theatre squares because I know something coming up later.

And for 6, there's also a lot of harder decisions. How fast do I need to push for my next governor title? Do I need a new governor, or just take the next level on one I have unlocked? Should I shift around my governors? I might "pre-move" Reyna to a city knowing that in 5 turns when they are set up, I will have a new title to unlock the double-harbor bonus, and at the same time, the harbor I'm building in the city will be finished. The ability to "look ahead" in the game is something that is a lot easier for humans to see the picture of than to figure out in a computer.

Never mind, I'm sure in a lot of ways, the hardest part of AI for a game like civ 6 is that the game is not "set", the game is eminently modable. It's a small thing, but consider that I could make changes to the files to put the spaceport as a district that you unlock in the culture tree. The game AI has to be able to handle that, and somehow handle that. How does the AI play if GDR combat strength is 20 higher than it is now, or 20 lower? I mean, I'm sure the in-game AI makes some assumptions (if unit A upgrades to unit B, unit B is better than unit A), but there's so many factors in the game that change that I know must be nearly impossible for the AI to reason. Never mind having the AI work for Maori/Babylon/etc... and every other civ that has some really weird individual bonuses, without having a custom AI written for each civ individually.

Sure, there's plenty of things the AI can do better than humans. Every time they unlock a policy card, it takes them a fraction of a second to evaluate every combo of policy cards and with some simple weights, optimize them. But even with that, you get the human strategy of "well I'm not building any melee units now, but I know my next policy unlock isn't for 10 turns, and I will probably switch a couple cities over to doing that soon, so I probably should slot in that card now."

 

[GeneralZIft]

If I have to keep reading 'ChatGPT AI' I might have to ask ChatGPT what would be the quickest way to jump off a bridge