Why Civ 6 AI is terrible - Part 1: Wonders

[Zike]

Disappointed by how terrible the AI is, I decided to make a series highlighting some of the flaws.

 

[Mojo85]

Good video. It’s clear the thought process is drown the player in bonuses to overshadow the lack of strategy

 

[kaspergm]

I don't disagree with your points about the AI in any manner, but I do think that's it's worth mentioning that if you have a huge production, and assuming you don't have anything better to build, building a wonder like Chichen Itza or Petra, even if you will get practically no yields from it, can be a valid strategy with the sole purpose to inhibit your opponent of building it. Petra in particular can be a literal game changer if you get it under the right circumstances, so given the circumstances mention before, it may not actually be a terrible idea to steal it if you can. Of course, if you're in the midst of an invasion, obviously these circumstances do not apply.

 

[Haig]

Well the combat AI is now better than in Civ 5 so that's pretty sweet. (Yeah yeah Civ 5 combat AI isn't great but..)

But of course a big shocking headline creates more views on youtube.

I say that for most of players AI is _good enough_.
Next I think they should concentrate on getting it to build more air power.

 

[Mojo85]

Higher difficulties should be harder AI, not just higher AI bonuses but the same strategy

 

[dexters]

This is always the case with Civ. Higher difficulty = production bonuses for the AI

Essentially Prince players and below get to pick almost any wonder they want to build.
Emperor and above players basically lose almost every early wonder they want unless they get a GP to rush build a wonder or is playing a specific Civ with wonder bonuses. They maybe able to nab late game wonders due to tech lead and other considerations but that's the limit to it.

That is part of the difficulty.

 

[Disgustipated]

Those wonders do make them hard to stop on cultural victories. Even in my King game, Scotland would get a cultural victory if I did nothing. I was forced to build a ton of theater squares to counter balance all the wonders and other things he has. Yeah it's still not a good use of production, but it does force the human player to do something about it often times if they can build enough wonders to threaten cultural victory.

How are you guys seeing the AI with St. Basil's Cathedral? Even though it can be built in all cities, so far I've only seen the AI build them in cities with Tundra, which actually surprised me. I've only seen it built twice though.

Now the second part I wholeheartedly agree. The AI should build nothing but units when in a war, that would make them dangerous, or at least not pathetic. It would be nice to code the AI to always build units (if they can afford them) during war. Though think about it, the United States built the Pentagon during WW2, we should have been building units. Of course we were never in any danger.

As for the last part, are you sure the AI can see that many tiles into your civilization? I'm not sure that they can. Yeah, I looked at it again, you can only see 2 tiles into his civilization, and you expect him to see 4 tiles into your civilization and see Big Ben half built? Unless he has a spy there, or a trader, I'm not sure he can see that. That cavalry up there I think can only see 2 tiles.

 

[Kwami]

Higher difficulties should be harder AI, not just higher AI bonuses but the same strategy

That's a nice idea, but it's not really practical to implement. Better strategy means longer turn times. It also means spending a lot of development resources to tune a more strategic AI that only a small minority of players will benefit from. Besides all of that, making a good, strategic AI is actually quite a bit harder than just piling on bonuses. Just about every 4X game cheats like this. It's simply a practical, non-ideal solution to a very difficult problem.

 

[Acken]

Higher difficulties should be harder AI, not just higher AI bonuses but the same strategy

Ideally but you'd have to start by having a good AI to begin with to be able to restrain its intelligence. Lacking that possibility, bonuses are fine. But the base AI still has to be decent so that you don't rely solely on bonuses. Difficulty is supposed to come from the combination of a capable AI with its bonuses. The fact that civ6 gives 2 free settlers on deity instead of one in civ5 and people are still able to win in 150 turns says a lot.

 

[hdbhdbvGFG541]

This is no surprise,civ 5 and 6 ai philosphy is :
"If it can build it,then build it.Regardless if it gives any benefits."
I mean how many civ v games did you see the ai building petra for a 1 desert tile.

 

[Siesta Guru]

Having done modding on civs AI building system, it's definitely a little more complex than 'build whatever I can', but the general gist of your video is correct in that it ignored about 95% of relevant info.

What it does is basically giving a score to every building it can build, and to units requested by 'contracts' (how that works is kind of irrelevant here).
It determines these scores based mostly on the resource yields + some extra factors it expects to get out of it. Civs can have differing desires for culture,food, etc. And it just multiplies that desire by the amount of gains.
Additionally, it can gain some bonusses/reductions to the score based on what kind of object it is, wonders get a different bonus from buildings/districts.
then it just selects the highest score, and builds that.

Unfortunately, the system is very underdeveloped. It for example generally doesn't care at all about 'complicated' effects, basically anything that isn't just a flat culture bonus or whatever. Chicen itza's rainforest effect is such a complicated effect, so it doesn't care at all. Because it doesn't understand these effects (present on most wonders) at all, and because it'd be weird if AIs never built any wonders, wonders are by default desirable enough that they'll sometimes be build, regardless of what its effect does. Some civs are also more likely to build certain wonders, again regardless of what the actual effects are.

And yes, the scoring mechanisms miss a lot of other important factors too, like being at war / being under siege, being locked out of producing stuff, the worth of the tile a wonder takes up, and so on.

As for why that is the case. While I would've done things differently, I can at least somewhat sympathize with them on this point. They (likely not the AI dev) chose to make the game very moddable, not just by modders but also by themselves (90% of the new expansion is a mod). They seem to have not hardcoded anything about buildings, policies, techs, and so on inside of their codebase. So most of the input the AI takes has to come out of the moddable xml files they made. And while taking the plain resource yields off of buildings and basing decisions based on that is pretty straightforward, and while dealing with just chicen itza would have been relatively straightforward, writing a generalized AI that is able to infer what a wonder does and then makes reasonable decisions on it is not so easy.

 

[blackcatatonic]

How are you guys seeing the AI with St. Basil's Cathedral? Even though it can be built in all cities, so far I've only seen the AI build them in cities with Tundra, which actually surprised me. I've only seen it built twice though.

Tamar built it in a desert city in my last game. I wasn't best pleased with her.

 

[Ryika]

There are many flaws with the AI, but building priorities, especially with wonders, are the one thing it doesn't really "need" to be good at, precisely because you can just add production bonuses to make up for bad decisions.

Wonders in particular are something that AIs SHOULD want to build even if it's not optimal, simply to have the "wonder race" be a thing. If anything, the problem is that these wonders are just too specific.

<And then add my usual comment here about how Firaxis does not seem to design things with AIs in mind, but instead designs things and then the AI guy has to work with what he or she gets presented with>

 

[UWHabs]

Having done modding on civs AI building system, it's definitely a little more complex than 'build whatever I can', but the general gist of your video is correct in that it ignored about 95% of relevant info.

What it does is basically giving a score to every building it can build, and to units requested by 'contracts' (how that works is kind of irrelevant here).
It determines these scores based mostly on the resource yields + some extra factors it expects to get out of it. Civs can have differing desires for culture,food, etc. And it just multiplies that desire by the amount of gains.
Additionally, it can gain some bonusses/reductions to the score based on what kind of object it is, wonders get a different bonus from buildings/districts.
then it just selects the highest score, and builds that.

Unfortunately, the system is very underdeveloped. It for example generally doesn't care at all about 'complicated' effects, basically anything that isn't just a flat culture bonus or whatever. Chicen itza's rainforest effect is such a complicated effect, so it doesn't care at all. Because it doesn't understand these effects (present on most wonders) at all, and because it'd be weird if AIs never built any wonders, wonders are by default desirable enough that they'll sometimes be build, regardless of what its effect does. Some civs are also more likely to build certain wonders, again regardless of what the actual effects are.

And yes, the scoring mechanisms miss a lot of other important factors too, like being at war / being under siege, being locked out of producing stuff, the worth of the tile a wonder takes up, and so on.

As for why that is the case. While I would've done things differently, I can at least somewhat sympathize with them on this point. They (likely not the AI dev) chose to make the game very moddable, not just by modders but also by themselves (90% of the new expansion is a mod). They seem to have not hardcoded anything about buildings, policies, techs, and so on inside of their codebase. So most of the input the AI takes has to come out of the moddable xml files they made. And while taking the plain resource yields off of buildings and basing decisions based on that is pretty straightforward, and while dealing with just chicen itza would have been relatively straightforward, writing a generalized AI that is able to infer what a wonder does and then makes reasonable decisions on it is not so easy.

Yeah, I can certainly understand how a moddable game can certainly wreck havoc with the AI. And as mentioned, on higher levels they have the production, so in some cases, you can make a case that building them for the era score along might make sense (except when they're being invaded, obviously). But yeah, especially with now 3 tile-specific wonders, it would certainly be nice if they could build in some notion of their value to try to only build them if they're fairly valuable. Like, there's almost enough of them that they could form their own "class" of wonder that should have its own tag in xml files, for example.

 

[Siesta Guru]

Yeah, I can certainly understand how a moddable game can certainly wreck havoc with the AI. And as mentioned, on higher levels they have the production, so in some cases, you can make a case that building them for the era score along might make sense (except when they're being invaded, obviously). But yeah, especially with now 3 tile-specific wonders, it would certainly be nice if they could build in some notion of their value to try to only build them if they're fairly valuable. Like, there's almost enough of them that they could form their own "class" of wonder that should have its own tag in xml files, for example.

Yes, this is what I would've done too. Either allow more complex ways of doing AI valuations in mods (more AI-specific fields, or perhaps through allowing lua access to change valuations), or build a system that allows both a simple valuation for simple and/or modded buildings as well as more complex hardcoded internal valuation systems for all the official game buildings. Most players will not be playing with heavily modded games anyway.
It's certainly not an unfixable problem, but given the way businesses work, someone up the chain probably just decided that this one was not worth the money/time to properly fix.

 

[Gort]

The AI builds space victory buildings in games with the space victory disabled

It uses policies with zero benefit to itself, like Colonial Taxes (+25% gold in cities not on your original continent) when it has no cities not on its original continent

 

[Siesta Guru]

Oh yeah the policies choices are abysmal. There's no way whatsoever to influence which they pick with mods, and the picks they do make are nigh indistinguishable from random picks on first glance. It isn't though, they're consistently terrible. There's a log file with ratings for every policy. These ratings are bizarre and appear to actually select policies in a way that's frequently actually worse than random. Basically limes every day all day in the early game, logistics in the late game. And apparently with R&F Civil Prestige has also become really popular (even before they have any 3 upgrade governors). Niche picks for us, bread & butter for the AI

Edit: haha, I just took another glance at it. In a game with absolutely nothing going on, here's a selection of policy ratings:
Charismatic Leader 21
Natural Philosophy 19
Scripture 76
Simultaneum 1018
Religious Orders 5192
Invention 32

Clearly, religious orders is 273 times as good as Natural Philosophy despite no religious wars going on. And obviously Simultaneum is more than 10 times as good as Scripture, even though the civ in question has no pop 10 cities while it does have adjacency bonuses.

 

[PhilBowles]

Well the combat AI is now better than in Civ 5 so that's pretty sweet. (Yeah yeah Civ 5 combat AI isn't great but..)

It's still worse than Civ V in its final form, and mostly worse than Civ V in Gods & Kings (which could, for instance, coordinate melee and ranged attacks, use an airforce and nuclear weapons, avoid standing within range of walls with units near death, use more appropriate unit mixes, attack cities when it has the forces to take them, and focus on attacking particular unit types with the units that counter them - all things Civ VI AIs are entirely incapable to at best inconsistently capable of).

The only ways in which it's consistently better than Civ V at any stage are that it can move and shoot with units (which was fixed in Civ V - I can't recall quite when but later than it should have been, maybe as late as BNW) and that it doesn't leave units embarked in sight of ranged units (something never fully fixed in Civ V). It does do the reverse, though - not recognise naval ranged units as a threat and so keep its units standing within range of them.

And all that said, Civ V's combat AI was always notoriously inadequate.

But of course a big shocking headline creates more views on youtube.

This post had nothing to do with combat AI. Building AI, despite widespread perception to the contrary, is more important. The AI literally can't win the game by domination however much the combat AI is buffed so it needs to be capable of competing peacefully, and compensating for poor combat AI by spamming units is a better fix than compensating for wasteful production choices with extra bonuses.

I say that for most of players AI is _good enough_.
Next I think they should concentrate on getting it to build more air power.

It now seems to build air units, it just can't use them. In my last game I saw AI bombers for the first time - Arabia built three of them. They never left the hanger.

Having done modding on civs AI building system, it's definitely a little more complex than 'build whatever I can', but the general gist of your video is correct in that it ignored about 95% of relevant info.

What it does is basically giving a score to every building it can build, and to units requested by 'contracts' (how that works is kind of irrelevant here).
It determines these scores based mostly on the resource yields + some extra factors it expects to get out of it. Civs can have differing desires for culture,food, etc. And it just multiplies that desire by the amount of gains.

That seems a basic system that could be easily tailored to change the priority of Wonders like Petra or Chichen Itza based on how many of the appropriate tile type it has access to - its score starts low and every tile of the appropriate type boosts it further up the priority list.

There are many flaws with the AI, but building priorities, especially with wonders, are the one thing it doesn't really "need" to be good at, precisely because you can just add production bonuses to make up for bad decisions.

The point of production bonuses is to make the AI a difficult opponent, not to compensate for its weaknesses. It already needs production buffs to compensates for the extra unit losses it takes as a result of being poor in combat, but that's much easier to compensate for than producing the wrong thing. If the AI never gets to a point where it's producing things that advance its victory progress, it's irrelevant how big its production bonuses are. Giving it an extra 700 because you anticipate it building Chichen Itza is irrelevant if it just ploughs the excess into a one-flat-desert-tile Petra.

I regret taking Dumbarton as aggressively as I did in my current game, now. The city was halfway through building Kokotu-In - I was keen to take it quickly in case it shifted production temporarily to walls, but of course it didn't.

 

[steveg700]

Glad to see this.

Be sure to send links to professional game reviewers.

 

[SharTeel]

For me, the worst part of the AI is the tendency to found cities 1 tile away from rivers/lakes. I mean, it's "fine" once you unlock aqueducts, but those extra settlers on emperor+ always end up in the worst locations...