[GS] New patch: AI walled city attack experiment results

[Anno]

It's funny because in the ~600 hours I have logged in Civ I think I've taken maybe 5 cities at most through war. Instead I mostly do what the AI did in the first series of tests, which is kinda knock some things around hoping that they are tempted into offering peace so I can go back to making the map look pretty.

 

[Architect]

See this thread is amazing because it shows how important the tests are that we run to try and understand the problems the AI has. The OP was trying to prove that the tactical AI can't attack and take walled cities. His test was flawed because the operational strategy AI was not to the point of being "angry" enough at the player to want to kill it. The second tests by V.Soma prove that an "angry" AI can take a walled city with no problem if its got the right operations queued up.

Maybe this is why the AI feels so poor some of the time. Things we all think should be easy for it just don't happen correctly and other times its just a wrecking ball. Not because it can't handle 1-UPT but because its operational strategy AI is confused or has too many things its trying to accomplish. Its like the second set of tests helped the AI focus on killing the player while the first set of tests the AI just thinks the player is not worth messing with.

Also, Carl the QA guy made a comment in the multiplayer stream that I found interesting. He said if you don't share your capital location with the AI it can't create a military operation to take it. I wonder if this test would change completely if you refused to share the location of your capital with the AI. There could also be a bug with the AI finding capitals with their vision versus finding capitals by exchanging that information with the AI in the diplomacy scene.

I've written AI for a CCG and the Civ AI is really quite good. I can promise you it is very hard to figure out why the AI is not working right. The better our testing becomes the more we'll help them.

 

[Casworon]

I think its definitely better than release. One of the things im most excited about with the possibility of a third expansion is another year of work on the AI

Above all that though DLL release would make me very happy

 

[Trav'ling Canuck]

I've observed the AI capture walled cities many times during my R&F tests, so I'm not surprised by @V. Soma 's results. I'm glad to see those results laid out like this.

I do believe that one challenge the AI has is that it's been taught that it needs to include Catapults/Artillery as part of the Operations that @Infixo referenced. While unlike on vanilla release, the AI is no longer stymied by walled cities, it's still not as effective at taking those cities as I believe it could be, and I believe this is related to the programming to include siege equipment in the City Attack Operation. This appears to me to have three negative effects on the AI's effectiveness on attack:
a) it slows down the AI's advance, as the attack won't proceed until the catapults/artillery are in position
b) it dissuades the AI from pressing the attack when the siege units are shot up (unless it's already damaged the city enough that it recognizes it can press forward and take it)
c) the AI only rarely (ever?) proceeds with ram-based city attacks, which are far more effective

Presumably the GS buff to walls will hamper the AI's city-taking abilities even more, until/unless it is programmed to avoid catapults and bring rams. This is a hypothesis only at this stage.

 

[acluewithout]

I've observed the AI capture walled cities many times during my R&F tests, so I'm not surprised by @V. Soma 's results. I'm glad to see those results laid out like this.

I do believe that one challenge the AI has is that it's been taught that it needs to include Catapults/Artillery as part of the Operations that @Infixo referenced. While unlike on vanilla release, the AI is no longer stymied by walled cities, it's still not as effective at taking those cities as I believe it could be, and I believe this is related to the programming to include siege equipment in the City Attack Operation. This appears to me to have three negative effects on the AI's effectiveness on attack:
a) it slows down the AI's advance, as the attack won't proceed until the catapults/artillery are in position
b) it dissuades the AI from pressing the attack when the siege units are shot up (unless it's already damaged the city enough that it recognizes it can press forward and take it)
c) the AI only rarely (ever?) proceeds with ram-based city attacks, which are far more effective

Presumably the GS buff to walls will hamper the AI's city-taking abilities even more, until/unless it is programmed to avoid catapults and bring rams. This is a hypothesis only at this stage.

Which is why they should buff Seige and nerf Rams. The AI uses the first nit the second, so instead of reprogramming the AI, just change the game rules to fit the AI’s strategy.

 

[Infixo]

c) the AI only rarely (ever?) proceeds with ram-based city attacks, which are far more effective

Which is why they should buff Seige and nerf Rams. The AI uses the first nit the second, so instead of reprogramming the AI, just change the game rules to fit the AI’s strategy.

Actually, the rare usage of rams vs. siege units has nothing to do with their strength, etc. and buffing or nerfing won't solve it. Battering Ram and Siege Tower are not siege units - they are support units. AI has a separate group called "siege support units" that includes also engineers, medics, drones, mobile sams, etc. For a city attack AI needs at least 1 siege or siege support unit. So, it sometimes produces a ram but the moment it gets to Catapults, it always takes them because they have a non-zero CS and rams have 0 because they are support units, not "normal" siege.

This puzzle is solvable. In my Real Strategy mod I have tweaked the AITypes for rams and AI uses them on a regular basis.

 

[Trav'ling Canuck]

Actually, the rare usage of rams vs. siege units has nothing to do with their strength, etc. and buffing or nerfing won't solve it.

I believe that @acluewithout 's point was that given the AI's preference for catapults, the reasons for which you've clarified - thank you! - it would be easier to nerf the impact rams have on walls and / or buff the bonuses catapults receive versus walls. Then the AI's preference for catapults will be less of a disadvantage for them vis-à-vis players.

However, it seems you've already solved the problem a different way, so that works too!

 

[679x]

I was one of the people that thought the AI was bad. This was an awesome thread to read and it's made me change my mind (somewhat). With Soma's tests, it now seems that the AI has trouble conquering cities because it really doesn't even want to conquer them in the first place, perhaps because of the diplomatic penalties it would incur? Of course, in a game with just 2 players, the one that does the conquering would suffer no diplomatic consequences, and would flat-out win the game. Maybe code adjustments can be made to alert the AI that it's basically guaranteed to win if it just conquers this city.

I wonder if this sort of problem comes up when the AI is presented with opportunities in other victory types besides domination? Maybe the AI is slow to do things such as win science victories because it's preoccupied with other things, and doesn't recognize the value in obtaining a victory? "Well, I just have to finish this last Mars reactor and victory is mine, but that city-state over there is getting annoying, so maybe I'll mobilize and go try to take it over instead. After that, I should focus more on finishing all the buildings in my Theatre Squares, because I'm falling behind on culture".

 

[DJ_Tanner]

I thought I'd re-do my traditional AI test to see if the AI can beat an opponent who doesn't fight back. As usual, we'll use the following settings to stack the deck as much in favour of the AI opponent as possible:

* Domination victory only
* Duel size map
* Pangaea
* Information era start
* Deity difficulty
* Online speed
* Aztec opponent
* No turn limit

I'll be restricting myself in the following ways:

* Delete all starting units except for rangers and one settler, found only one city
* Declare war immediately upon meeting the opponent
* Do not build a military or fire with city attacks

I think your settings are directly leading to your results. The AI, afterall, is simply a decision tree. So it must categorize and rank the options available to itself. By removing any threat whatsoever you are basically giving the AI no reason to rank the decisions to remove you high enough for it to actually engage in that properly.

 

[Infixo]

The AI in Civ6 (and in Civ5) is actually a few systems working independently most of the time. Tactics, strategy, ops, diplo, wc, etc. There is unfortunately too little interaction among them and this is what most people perceive as „bad AI” usually, imho. Human brain sees connections and interactions everywhere, Civ6 is just a computer program. Those need to be programmed and configured.
Simply speaking - AI doesn’t „see” a goal and doesn’t plan how to get there. It makes hunderds of small decisions every turn and they eventually add up to a victory. Those decisions are usually based on scoring or valuation, and this is the place where some „goals” or strategies can be applied. All you can „teach” AI is e.g. „Build more campuses if you go for science” or „build more miltary if you are agressive”, etc.

 

[Acken]

The AI in Civ6 (and in Civ5) is actually a few systems working independently most of the time. Tactics, strategy, ops, diplo, wc, etc. There is unfortunately too little interaction among them and this is what most people perceive as „bad AI” usually, imho. Human brain sees connections and interactions everywhere, Civ6 is just a computer program. Those need to be programmed and configured.
Simply speaking - AI doesn’t „see” a goal and doesn’t plan how to get there. It makes hunderds of small decisions every turn and they eventually add up to a victory. Those decisions are usually based on scoring or valuation, and this is the place where some „goals” or strategies can be applied. All you can „teach” AI is e.g. „Build more campuses if you go for science” or „build more miltary if you are agressive”, etc.

While it is true that the AI code is individual parts it is not really why it is considered bad. Good coordination between all systems would be required to make a competitive AI but I think people at this stage just want some semblance of competency.
Especially in 1UPT combat since it's been 2 games with that system with no progress (3 if you count CivBE). Hell at least in Civ5 the AI could muster decent attacks above prince.

 

[Trav'ling Canuck]

The AI in Civ6 (and in Civ5) is actually a few systems working independently most of the time. Tactics, strategy, ops, diplo, wc, etc. There is unfortunately too little interaction among them and this is what most people perceive as „bad AI” usually, imho. Human brain sees connections and interactions everywhere, Civ6 is just a computer program. Those need to be programmed and configured.
Simply speaking - AI doesn’t „see” a goal and doesn’t plan how to get there. It makes hunderds of small decisions every turn and they eventually add up to a victory. Those decisions are usually based on scoring or valuation, and this is the place where some „goals” or strategies can be applied. All you can „teach” AI is e.g. „Build more campuses if you go for science” or „build more miltary if you are agressive”, etc.

Presumably there's a hierarchy, too, from a development perspective. You need to teach the AI to move units, teach it to build districts, etc., get all of the various base systems working first, before you can implement things like "now we're attacked, so we need to interrupt our normally scheduled build process to bring you this new priority".

 

[Eva-kun]

Which is why they should buff Seige and nerf Rams. The AI uses the first nit the second, so instead of reprogramming the AI, just change the game rules to fit the AI’s strategy.

Wish they could be program to create a formation with a ram for their melee units. They do use rams, but not all that well at the moment. Siege does need a buff. Perhaps taking an extremely low amount of damage from city attacks and range attacks? I find it funny when I do use siege units, they take a massive amount of damage from a city attack.

The usual problem is walls being able to attack and causing damage to melee attackers. As a human player, I tend to use range units/siege units to destroy the wall down or use melee unit with a ram to destroy the wall. Not constantly use my melee units to attack a city's health down. The AI will just bash their melee units, with or without a ram, into a city and take massive damage to their melee units or even die doing it. If they have tech advantage, it's not a problem, but when tech even, it's just sad to watch.

 

[V. Soma]

I will do test with my city far from AI and not sharing, still making AI declare on me

EDIT:

setting is same as in OP,
plus: my city is far from AI and I don’t share visibility, and still make AI declare on me, I will actively defend in my city

211: I settle really isolated coast place… ice-block to the west and east too
I keep a mech. gun for defence

215: I meet Monty’s ranger, some 10 tiles from me - I denounce - I go on to explore his city location

218: I find Monty city, his land is some 15-20 tiles from me - he has not seen my land yet

220: Monty sees my city, relationship is -16

227: Monty completes Earth Satellite

230: I went around Monty’s empire, the center is ca. 25 tiles from me - I deleted all exploring units

235: I denounce and demnad again, Monty has lots of resources, relationship is only -4 (luxury parity)

249: Aztec has new city - I cannot make Monty angry, relation is 0 now…

253: Monty switches to Fascism, finishes Moon Landing, would win Culture J

257: I denounce again (it has no effect on relation, it is now at -4)

260: Monty wants me to build ships - relation is at -11 now

267: Monty is happy to have more great people than me - - relation is at -5 now

270: Monty is now Corporate libert.

279: I denounce and demand again (Monty has 5 cities plus capital)

300: I denounce - relation is -3

311: sea level rises 1 m

315: Monty finishes Laser Station

OK I give up - Monty is not about to win this domination game…

 

[Takfloyd]

I'm wondering, at this point, if it's intentional. Are they catering to a wider audience by having a less challenging game? I mean, it is off-putting to dump 8 hours into a game and then get steamrolled by the computer. Maybe they're curtailing that frustration by making the combat AI intentionally dumb. Pulling its punches, so to speak.

Not combat AI, but has anyone noticed that the AI doesn't actively try to convert your cities in pursuit of a religious victory if you have your own founded religion? I've seen waves of proselytizers skip my civilization entirely.

Yes, that has been theorized before...ie, that 'The Masses' don't really want to LOSE the game but they want to feel like they 'beat' the game. I think there is some merit in that argument (especially given the massive early successes of games like Farmville and their ilk where mere participation was enough to 'win').

In a crowded market, I imagine they feel that many players simply want to go through the motions and then win the game at the end. If they play for 4 hours and then get steamrolled by an AI mega-stack, they'll rage quit and quite possible never come back.

Of course for many players the opposite is true...if they don't feel like they have been challenged, they lose interest and don't come back. But at least those players are far less likely to come to public forums and ratings system and blast the game, the devs, the dev's dogs, and the world at large.

Whether they are correct in thinking that players don't REALLY want to lose or not still remains to be seen. Given the sales for this game, I'm sure their not crying too much over the criticism of the AI.

The thing is, the game has something called a difficulty setting for this purpose. If you're playing on Deity, you aren't going to want the AI to pull its punches. Firaxis must understand this.

Also, Civ VI has sold far less than Civ V and has a much lower user rating on Steam, so I'm sure they've taken note of the criticism. During Rise & Fall when user satisfaction was at its lowest, they took active action to turn it around by being more transparent with their dev procress and adding requested features like Join Ongoing War. So it's absolutely worth it to keep pestering them about the AI.

 

[Krajzen]

Also, Civ VI has sold far less than Civ V and has a much lower user rating on Steam,

Has it? Where do you take info on sales from?
Regarding steam and meracritic ratings, not sure about you but I actually checked them for years and they were much worse in the past ane then gradually improved with G&K and BNW. There was a point when on metacritic civ5 had "mediocre" rating around 6.7, then I recall G&K having "good" rating around 7.5 and then BNW reached "very good" ratings around 8.5. Nowadays civ5 vanilla user rating on metacritic rose to 8.0, from mediocre to very good category, so I guess the same happened with steam scores over time.

 

[Haggbart]

Has it? Where do you take info on sales from?
Regarding steam and meracritic ratings, not sure about you but I actually checked them for years and they were much worse in the past ane then gradually improved with G&K and BNW. There was a point when on metacritic civ5 had "mediocre" rating around 6.7, then I recall G&K having "good" rating around 7.5 and then BNW reached "very good" ratings around 8.5. Nowadays civ5 vanilla user rating on metacritic rose to 8.0, from mediocre to very good category, so I guess the same happened with steam scores over time.

This x 1000.

I think people forget how incredibly poor Civ5 was at launch. A lot of the mechanics we take for granted today was not in (city state quests, religion, tourism, espionage etc). However it had some great ideas and huge potential, but I'm still shocked how good they managed to make the game with expansions and patches.

Civ 6 was a much better at launch, but of course it's harder to improve upon an already great game.

The AI has been been a point of contention, even before 1upt. But the new tests show that it's important to see the AI behaviour in context with circumstances in the game. It's not AI ability to take cities per se, but rather its willingness to take cities in situation X and Y. And those different circumstances also make direct comparison to other strategy games invalid.

 

[bbbt]

Also, Civ VI has sold far less than Civ V and has a much lower user rating on Steam, so I'm sure they've taken note of the criticism. During Rise & Fall when user satisfaction was at its lowest, they took active action to turn it around by being more transparent with their dev procress and adding requested features like Join Ongoing War. So it's absolutely worth it to keep pestering them about the AI.

Most of the new negative reviews on steam are about the Eula so I'm not sure theyd get much of a game related message from that.

 

[Krajzen]

This x 1000.

I think people forget how incredibly poor Civ5 was at launch. A lot of the mechanics we take for granted today was not in (city state quests, religion, tourism, espionage etc). However it had some great ideas and huge potential, but I'm still shocked how good they managed to make the game with expansions and patches.

Civ 6 was a much better at launch, but of course it's harder to improve upon an already great game.

The AI has been been a point of contention, even before 1upt. But the new tests show that it's important to see the AI behaviour in context with circumstances in the game. It's not AI ability to take cities per se, but rather its willingness to take cities in situation X and Y. And those different circumstances also make direct comparison to other strategy games invalid.

Civ5 is the biggest "mediocre game on release turned very good with time" case known to me personally.
Civ5 on release was catastrophic.
- Bugs, technical issues, bad performance
- Multiplayer borderline unplayable
- Crazy unbalanced (example: companion cavalry unit replaced already OP horseman at the same cost, was faster, generated generals at double speed and had 40% more strength - it was devastating)
- Ridiculous early mechanics, such as cavalry unit line ending on dead end in industrial era, or xbowmen upgrading to riflemen and completely wasting ranged promotions
- Extremely dysfunctional AI which was simultaneously completely incapable of any diplomacy and of conquering any cities
- Combat system scale 1-10hp instead of 1-100hp until G&K, with guaranteed 1hp dmg, which meant you could very reliably kill giant death robots with ancient archers
- No religion, espionage, trade routes, tourism, ideologies, world congress, city state quests, diplomacy etc.

This forum was such ****storm for few first months

EDIT
Jesus, I completely forgot there were no denouncements and friendship declarations in civ5 1.0, AI just randomly disliked you with no feedback and no possible reaction

 

[Leathaface]

I hope Firaxis read these thread and learn from it.