My Yearly "Is the AI Any Good Yet" Thread

[IMADAJ]

I’m still waiting on a deep mind AI like Alpha Star to be implemented into Civ and other games.

[spolier]Although there have been significant successes in video games such as Atari, Mario, Quake III Arena Capture the Flag, and Dota 2, until now, AI techniques have struggled to cope with the complexity of StarCraft. The best results were made possible by hand-crafting major elements of the system, imposing significant restrictions on the game rules, giving systems superhuman capabilities, or by playing on simplified maps. Even with these modifications, no system has come anywhere close to rivalling the skill of professional players. In contrast, AlphaStar plays the full game of StarCraft II, using a deep neural network that is trained directly from raw game data by supervised learning and reinforcement learning.[/spoiler]
https://deepmind.com/blog/article/alphastar-mastering-real-time-strategy-game-starcraft-ii

One day I’m hoping and expecting most video game AI to actually be able to better challenge humans without depending on extra bonuses and perks alone. I’m looking forward to the day I can play Civ single player and have it feel like a true multiplayer game with distinct individual rulers and different play styles depending on their civilization actively going for a win.

edit: I closed the page with quote. I’ll see if I can find it

 

[kaspergm]

I’m looking forward to the day I can play Civ single player and have it feel like a true multiplayer game ...

I know taking the quote out of context is a bit unfair, but I definitely see different groups of interests forming here. Personally, I don't want a game vs. AI to play out like a multiplayer game. I think when you play multiplayer (with emphasis on think, because I don't actually myself play multiplayer), the default focus will be a race towards victory. On the other hand, there's a much larger immersion element in games vs. the AI, at least from my point of view.

If you want the AI to be a challenge, you should program it to always focus directly and only on military and to rush the human player. That would make it very hard, if not impossible, for the human player to survive. But it wouldn't be that much fun as a gaming experience imo. I like the fact that the AI plays the game as it's intended to, not necessarily in the way that makes it win fastest. One could program the AI to spam campuses in every city, but that's not the way the game is intended to be played (as I see it). The idea of the game is that you should have a balanced mix of districts, and that's basically what the AI builds.

My main problem with the AI is that it performs so poorly at the task its intended to do. It's supposed to play the game as its intended to be played, but it does that poorly. It's gotten better at placing districts, but it's still bad at aspects like using policy cards and moving and applying military units. Particularly the latter is the big joker, and obviously an extremely hard problem to code. I wish the AI was better at playing the game, but I don't wish it would play it like a human player does. If that makes sense.

 

[IMADAJ]

Given the way that DeepMind teaches itself I don’t think we would see a pure military rush or campus in every city. Both of these play styles each have their own weakness, and DeepMind would have learned how to exploit those and thus would switch gears to compensate. Each AI controlled Civ would also be working towards its own win as well so there wouldn’t be an all out gang up on the player (although it could still play out that way).

I understand the immersion you spoke of and it’s one of those things that made me fall in love with the game long ago. I personally think better AI would drastically enhance that. (Personal opinion. I’m not saying mine is right or better than yours)

As an added side note I want to point out that difficulty could still be adjusted while using better AI.

 

[AriochIV]

I also don't want the AI to behave like a human player who is only trying to win; I want the AI to play as if it was a "real" civilization.

If you want an opponent that plays like a real human opponent, then play multiplayer. No AI opponent will play as well as a real human opponent. Not anytime soon, anyway.

I wish that Firaxis put more emphasis on coding better AI, but they know that the vast majority of their customers play only on the easiest difficulty settings. They really don't care whether the AI is a challenge for the hardcore players when they can just have it cheat harder on the higher difficulty levels.

 

[Tabarnak]

If you guys want the AI to be like humans don't forget to make them whining and insulting you when gaming.

 

[Solver]

I'm always saddened by blame being put sorely on the developer like above. It's not Firaxis that sets the game's budget, or the resource allocation. They don't get to determine the priority given to the AI, or modding, or voice acting, or graphics compared to one another.

Do I wish the AI had a higher priority? Absolutely. But that's not really a Firaxis-level decision, and the AI is quite impressive when you remember the majority of it is one person's work.

Also sadly, I don't really see a scenario where the AI would become a much larger priority in the future. Most people play on Prince or below. To satisfy the hardcore players like many here, it's not just a matter of letting the AI engineer work for a month longer. They'd need a dedicated AI engineering team of 3-4 people, and additional QA effort focused on the AI. So easily 4x or 5x the current AI budget, that's the minimum investment necessary to really make a difference for the hardcore "Deity isn't much of a challenge if you survive the initial rush" players here.

 

[Bactrian]

If you want the AI to be a challenge, you should program it to always focus directly and only on military and to rush the human player. That would make it very hard, if not impossible, for the human player to survive. But it wouldn't be that much fun as a gaming experience imo. I like the fact that the AI plays the game as it's intended to, not necessarily in the way that makes it win fastest. One could program the AI to spam campuses in every city, but that's not the way the game is intended to be played (as I see it). The idea of the game is that you should have a balanced mix of districts, and that's basically what the AI builds.

I think this reflects a fundamental problem with the design of the game; namely, that it tends to reward all-in strategies instead of balanced ones. There's no particular reason you couldn't make it so that military potency is dependent on having a strong economy, whether that's in production or in gold. Same for any other particular balance of science vs. culture, production vs. faith or what have you. I've mentioned elsewhere on this forum that I think that the multiple victory conditions is hampering Civ because it generally rewards players for pursuing them in a single-minded fashion from the beginning of the game. This is why I'm hopeful for Humankind, mentioned above, because I really, really want to see what the effect of having a unitary score-type victory has on the game. Maybe I'm wrong, but I think it will be very beneficial.

My main problem with the AI is that it performs so poorly at the task its intended to do. It's supposed to play the game as its intended to be played, but it does that poorly. It's gotten better at placing districts, but it's still bad at aspects like using policy cards and moving and applying military units. Particularly the latter is the big joker, and obviously an extremely hard problem to code. I wish the AI was better at playing the game, but I don't wish it would play it like a human player does. If that makes sense.

Yes, this 100%. I find it galling that the AI can't even do what it's designed to do properly.

I'm always saddened by blame being put sorely on the developer like above. It's not Firaxis that sets the game's budget, or the resource allocation. They don't get to determine the priority given to the AI, or modding, or voice acting, or graphics compared to one another.

Do I wish the AI had a higher priority? Absolutely. But that's not really a Firaxis-level decision, and the AI is quite impressive when you remember the majority of it is one person's work.

I mean, I guess? I think Firaxis is a good company doing good work but I don't have any personal loyalty to them so I won't hesitate to critique their products. It may very well be that what I'm critiquing is a 2K Games-level decision, but I honestly consider that kind of distinction to really be neither here nor there.

 

[kaspergm]

If you guys want the AI to be like humans don't forget to make them whining and insulting you when gaming.

Well, aren't the already? Isn't that basically what the agenda system boils down to?

 

[Pfeffersack]

The good or bad thing about Civ6' AI is that it would already greatly benefit from a couple of rather small fixes (like not sacrifing both military and civil units to visible threats, spamming spaceports, various trade exploits including joining wars for 1 Gold and being more focussed when connecting ressources to just to just name what immediately pops up in my head when thinking for less than a minute).

Aside from that, some rule changes could help the AI (e.g. make walls less powerful or make Amenities tighter, as the AI goes for luxuries and the happiness yielding districts). All that would still not create a super-AI of course, but it would clearly help them and help to keep the immersion (currently there is rarely a turn, where I don't stumble over a serious AI plunder)

The latter would also benefit from a more varying behaviour of the AIs. I wish to have sharpened personalities, where one AI clearly goes cultural, while others spam holy sites or units. All AIs behaving the same is not good, not even if they do it well - e.g. I don't like that any AI, even leaders with environmental agendas, destroys the planet in the same way by mindlessly chopping down everything and building coal plants. In my ideal game, the heavy industrialist should do exactly that and the other civs to a lesser extend...but the environmental agendas should lead to civs just not willing to build coal plants, instead preserving forests for National Parks. Yes, I'm aware that this would make them less strong on the production front, but overall the game would benefit, if some AI play less than "optimal" in terms of winning (in some victory conditions). Even if that means that such a civ will get eaten in 9 out of 10 cases by an aggressive industrial neighbour - in the 10th case they might be a contender for Cultural victoy and in the rest of the games, they might be the fuel for the aggresor to become big enough to go for a world conquest.

 

[oSiyeza]

Tactical AI has improved a lot from vanilla but strategic AI is still very lacking. Does not pursue victory, doesn't try to win, just "hangs out". Usually too easy to friend and stayed allied/friends with the entire game.

All that said I still enjoy playing the game, it's just a different kind of game now so gotta set expectations real

I have tried AI mods but they have their own flaws. Eg Real Stategy tries to fix the strategic short comings but I have tried it multiple times and what ends up happening is AI goes for early religion a lot less often and makes RV too easy

Which is one of the reasons i like the mod. RV, is just a boring spam unit mess. The less AI civs waste their resources on that, the better.

 

[AsH2]

Well, aren't the already? Isn't that basically what the agenda system boils down to?

Next step is to make them rage quit.

 

[NukeAJS]

I don't know of ANY current 4X titles that will impress you with their AI.

Stellaris has good AI. Partly that's because its combat is much simpler. That doesn't mean its combat is less rewarding or fun. It just means that the way they setup the combat system is easier for an AI to grasp. If you haven't played or watched, combat is basically -- make fleet, fly fleet to enemy fleet, start "rolling the dice". The only variables are: number of ships, types and levels of defensive/offensive tech, size of ships, position of fleets just before engagement. The AI is able to manage all of these, except the last one, on the fly. It will usually avoid fleets it can't match. It will retrofit equipment based on what you're using. It will build ship types that counter what you're using. The only thing it can't do is warp directly into an enemy fleet on purpose. It's only really useful if you've a bunch of small ships and they've large ships but it's so useful in that situation that it's really the only way you can hope a bunch of corvettes can actually win a battle against several battle cruisers instead of just run interference.

Where Stellaris really shines is its strategic layer. The AI plays to win and it will actively try to stop the leader if it thinks it can somewhat effectively do that without losing too much. Another trap they avoided is historical accuracy. Most people know who Alexander the Great was and what he did. So, when he doesn't attack people, people criticize the AI. It's vice-versa with Gandhi. Since all of Stellarises factions are fictional, there aren't expectations. Even then, the ethos of a faction is almost always in-line with their actions.

 

[kaspergm]

Isn't there some reason to believe that a combat system like the one Humankind uses (from what I understand at least) will help the AI a lot? From what I recall, it sounded very much like the way games like Heroes of Might and Magic used to work, i.e. you group different units into an army, and the combat plays out on a tactical map. Heroes of Might and Magic had an AI that was pretty capable of handling such a system 20 years ago, so one would think it was doable these days also.

 

[Bactrian]

Isn't there some reason to believe that a combat system like the one Humankind uses (from what I understand at least) will help the AI a lot? From what I recall, it sounded very much like the way games like Heroes of Might and Magic used to work, i.e. you group different units into an army, and the combat plays out on a tactical map. Heroes of Might and Magic had an AI that was pretty capable of handling such a system 20 years ago, so one would think it was doable these days also.

That's true, but 1upt has been a staple of wargames for two or three decades, with many possessing an AI ranging between competent and deadly, and we all know how that turned out in Civ. Just because someone else has done a good job doesn't mean Firaxis will.

 

[Abaxial]

In my last game Norway attacked Yerevan, enabling me to decalre protectorate war. But I was interested to see that Yerevan made a fairly decent attempt to take a Norwegian city on its own, sending out a balanced army against it. It might even have taken it, but someone else took the suzerainty away, and all the CS units were recalled, even if I got suzerainship back instantly.

The main problem with AI combat generally seems to be its habit of sending units into battle piecemeal into the face of ranged fire. A simple rule like "don't advance into fire zones unless you have artillery support" would improve matters a lot.

 

[IMADAJ]

There is always the way Civ 1 was. Units were stackable, but if the stack was attacked the unit with the highest defense was used. If it died then the entire stack was destroyed.

 

[Bactrian]

There is always the way Civ 1 was. Units were stackable, but if the stack was attacked the unit with the highest defense was used. If it died then the entire stack was destroyed.

I kind of love the absurd results that you could get in Civ I. I remember raging out of mind (hey, I was a kid) when my battleship would occasionally die attacking a settler unit. The civil war mechanic that could split huge empires if you took their capital was also great.

 

[Linklite]

Right, that's definitely correct, I forgot. I see them with Musketmen and no Niter all the time. They do seem not to have cavalry though, at least in my experience, going for the cheaper units.

This touches pretty close to a question that came to mind at the beginning of the thread.

Do the Civ AIs actually use resources? Or do they just churn out units? I know old strategy games used the latter tactic, but I was curious about this one.

 

[oSiyeza]

This touches pretty close to a question that came to mind at the beginning of the thread.

Do the Civ AIs actually use resources? Or do they just churn out units? I know old strategy games used the latter tactic, but I was curious about this one.

The devs said in numerous occasions that the AI works with the same info and the same rules than the human players. With the exception of the difficultly bonus.

This is a very big design mistake in my opinion. The player will never play the same game of the AI. From the moment the player is able to gather knowledge from different matches, and is able to analyze design patterns, long term strategies, and information from previous games that include the AI behavior itself (what we call meta gameplay). The human player has access to game information, design rules and strategies that the AI is simply designed to not have.

Another terrible flaw is the over emphasis on designing the game to be as fair and multiplayer friendly as possible. Fxs clearly intended to avoid at all cost an snowballing AI that could frustrate or take away the agency of the human player. Not only we have as a result the first AI in the history of civilization that cannot and will not win a Domination Victory. Many other mechanics seem to be implemented with this principle in mind as well. As nothing should unbalance the game, many systems are just bland and pointless. Thus, the AI feels pasive, lacking in personality and more a decorative element than an active agent. Even more; as all the interactions with the AI feel artificial and passive, the human player fails in attributing a mind to the AI, so we are even unable to perceive the things that the AI may manage efficiently.

Finally the last big design mistake that handicapped the AI from the beggining is making the AI ignore the game civ is, to play a sim game that civ is not. The AI has not been programmed to win, but to act as a nice character and to use all the game systems. The AI is programed to roleplay, and to roleplay from the perspective of Fxs; not uppseting the player and using all the sytems and wasted mechanics that Fxs wants the player to use. But many of those systems are counterproductive to pursue the two main goals of the player; to have fun and to win.

I think other issues with the AI are also a result of rushed programing, not enough resources and a disconnect from the people that was adding mechanics to the people (more likely the one guy) that had to program the AI to use them.

No matter how many patches you add on top, these are issues that lie at the very core of the design and are very difficult to fix. I have to say however, that at least they have been trying, and hope they keep doing it.

 

[IMADAJ]

The devs said in numerous occasions that the AI works with the same info and the same rules than the human players. With the exception of the difficultly bonus.

This is a very big design mistake in my opinion. The player will never play the same game of the AI. From the moment the player is able to gather knowledge from different matches, and is able to analyze design patterns, long term strategies, and information from previous games that include the AI behavior itself (what we call meta gameplay). The human player has access to game information, design rules and strategies that the AI is simply designed to not have.

Another terrible flaw is the over emphasis on designing the game to be as fair and multiplayer friendly as possible. Fxs clearly intended to avoid at all cost an snowballing AI that could frustrate or take away the agency of the human player. Not only we have as a result the first AI in the history of civilization that cannot and will not win a Domination Victory. Many other mechanics seem to be implemented with this principle in mind as well. As nothing should unbalance the game, many systems are just bland and pointless. Thus, the AI feels pasive, lacking in personality and more a decorative element than an active agent. Even more; as all the interactions with the AI feel artificial and passive, the human player fails in attributing a mind to the AI, so we are even unable to perceive the things that the AI may manage efficiently.

Finally the last big design mistake that handicapped the AI from the beggining is making the AI ignore the game civ is, to play a sim game that civ is not. The AI has not been programmed to win, but to act as a nice character and to use all the game systems. The AI is programed to roleplay, and to roleplay from the perspective of Fxs; not uppseting the player and using all the sytems and wasted mechanics that Fxs wants the player to use. But many of those systems are counterproductive to pursue the two main goals of the player; to have fun and to win.

I think other issues with the AI are also a result of rushed programing, not enough resources and a disconnect from the people that was adding mechanics to the people (more likely the one guy) that had to program the AI to use them.

No matter how many patches you add on top, these are issues that lie at the very core of the design and are very difficult to fix. I have to say however, that at least they have been trying, and hope they keep doing it.

And thus the use of Deepmind and Alpha Star. An AI that has taught itself how to play using the cards delt per game, has the learned ability to adapt, even feint, and would recognize if one Civ was about to snowball and adjust via alliances to prevent it before it happened.

It learns from playing more games than a human could play in a lifetime in a matter of weeks. In previous examples Alpha Star has taught itself advanced strategies and techniques that took humans 10 or 20 years to come up with all in its own.

It’s not about difficulty either. AI can be made to make mistakes to limit how good it is so difficulty could be adjusted. It’s about playing against a believable opponent.