Has game AI stagnated?

OK, so yes, the cop cars in GTAV now do a blocking move, but is that really 12 years better than their moves in III?

And the enemy in the latest FPS? A bunch of guys run out and take cover, peek their heads, then exactly on time lean over their cover to shoot? Maybe run back and forth between cover? Didn't we see that playing Manon Batiste in Underground?

(I got in an argument about this somewhere else and thought I'd take a beating here, too.)

No it hasn't stagnated. The issue is that "AI Is Hard". We need more than 12 years to fix the problem.

I personally don't think it's a focus, at least in the action game arena. However, if you look at a game like Crusader Kings (I assume the same is true for EU) , the AI manages to be both unpredictable and challenging.

So I think we can make pretty good AI these days, but what you're talking about is a shooter, where the objective of the designer is to give you a fast-moving experience where you're constantly moving forward. If they go and make the enemies too difficult, that'd just slow you down and keep you from seeing the next exploding helicopter.

It's not just that AI is hard to do, but also I think that in AAA games they have such tight schedules, very little margin for error and so god damn much to do that they end up using whatever works for whatever kind of game they are making, in other words they play it safe.

I believe that if we ever see a real fantastic breakthrough in AI it will be on a small indie title, were they are not tied to publisher suits breathing on their necks have more creative freedom.

I would say yes it has, hopefully the new technology will create more advancement in AI programming.

Also GTA5 was kinda a bummer considering just get out of the car and hide behind a building works better than driving far away as fast as you can.

Let me be more clear: The issue with AI is that especially in games the capabilities expand in both ways which means it may never effectively catch up. It isn't that AI has stagnated or gotten worse but consumers have demanded more features in bigger scope which causes complexity to skyrocket making it hard if not impossible to create a capable AI around it.

To simply explain it, take a look at GTA and GTA5. The technology that can handle GTA compared to GTA5 has grown manifold but the amount "actions" any "actor" in GTA compared to GTA5 has grown even more. By the time GTA6 is out technology will have naturally grown but the amount of features and interactions will have gotten much bigger as well.

The best AI have rules and limitations that purposely constrain actions and reactions of not only the player but any "AI Actors" in the game. The current environment for game development is going for "unlimited actions and reactions" especially in newer and bigger designs like MMOs and MOBAs which promote play through open ended interactions. In this situation AI will never be as good as the features offered or even wanted by the player.

That's true, but whenever a question about good game AI comes up, FEAR still plays a large part in the discussion. You would think that someone, anyone, might have been able to make a first or third person shooter whose AI is more impressive than what came out in 2005.

EDIT: Nevermind, I jumped the gun and didn't see your later post. Still, it's a little disappointing that FEAR hasn't been beaten yet, even in the realm of linear first and third person shooters where the set of actions a character can perform is more limited than that in open world games. Besides, what have they really added since then? Cover systems? And maybe ziplines in the case of Tomb Raider?

And you end up with A) the goal for having better AI may have retired itself with online play .... and B) the definition of AI itself.

Borderlands' "randomization" of the weapons was a way of producing a varying experience on a next playthrough, but again, 10 guys attacked, 2 were big and tough, and a couple of them were suicide bombers.

MMOs and MOBAs benefit from the fact that the other characters are human, and therefore unpredictable. So with Titanfall and no campaign, the issue probably won't come up.

Still, graphics have gotten great over the last 10 years. Frame rate is increasing. And like a few of you said, the devs are not charged with making the enemy smarter, or adaptable, or inventive.

AI is bad in games like Call of Duty, but pretty good in games like Halo, Killzone, and Crysis. Stealth AI has always been stealth AI, take it or leave it. Borderlands AI is pretty bad but it sort of works in the way that it is bad relative to the gunplay. Defiance AI was alright.

The non FPS games with the "best" AI are heavily scripted, which works just fine.

I'm constantly underwhelmed by AI in videogames, which is probably why the majority of games I now play are entirely online. Fuck shiny graphics, we should measure the passing of a generation in the versatility of it's robot brains.

The problem is that whenever a programmer makes a quantum leap forward in AI technology, they are assassinated by a time traveler in order to prevent a dystopian future where humans must wage a war against emotionless robots controlled by a near-omniscient network hive mind.

I suppose. It doesn't seem like it's progressed much lately, although I know that's a complex problem without a simple solution. I guess that's why we've been seeing a push for integrated online feautures.

Dark Soul's AI is pretty intense.

Not really, it's just the nature of the animations. For instance, getting attacked while you heal is a combination of both certain enemies being aggressive enough to attack during the animation and the animation being long enough that you are likely still vulnerable when they reach you. Likewise with trying to doge and parry attacks. The animations are just a little faster or slower than you would think they would be and it messes up the timing. Really, the Dark Souls AI is pretty simple and exploitable when get down to it, like kiting multiple enemies out of a large group with a bow, like the Painted Guardians. If they had good AI, they would all just rush you as soon as you attacked one.

I would take it a step further and say that AI has regressed.

Eh? Single player shooter campaigns still have a lot of popularity. Tomb Raider, Halo, Call of Duty, Far Cry 3, The Last of Us, and Metro Last Light are all first or third person shooters with dedicated single player components. Time and money would not be put into a single player portion of a game if the publishers and developers didn't think it was worthwhile.

The definition of AI hasn't changed, it still has to do with how the computer reacts to a player's presence. None of the things you mentioned have anything to do with artificial intelligence except Borderlands, and that game does not have good AI.

As far as framerate goes, it seems like virtually every AAA release these days has framerate problems. If you go back to the PS2 era, you'll find that framerate problems aren't as egregious as they are in, say, Mass Effect. Or Far Cry 3. Or even Borderlands, which has a bit of slowdown pretty much every time you kill something.

Well I think a lot of the dev efforts have gone to keeping up with visual improvements, which require a lot of man hours. Not to mention that most FPS efforts these days go into the multiplayer aspects of the game due to used game sales etc. Single Player is not profitable for them.

That being said AI is light years better than the early Ps2 era, maybe not in the top titles, but you rarely see really really bad AI as it used to be very common place back then in mid and low end titles. Now that the graphics war feels like it's coming to a close I'm hopeful that improved Ai will become a focus once more.

AI has actually gotten worse. F.E.A.R had amazing AI and then the industry went downhill. I think design and creation as a whole in games (sans a few exceptions) has hit the rocks because people genuinely believe graphics=progress and that isn't true at all.

I'll tell you when Titanfall comes out

I'm more than willing to assume that many devs put more work into other aspects of a game than progressing AI, but there's another crucial point that's often left out of this conversation. A lot of gamers don't want that big of a challenge.

The goal isn't to make AI that's indecipherable from a human, or perhaps even better, because tons of players would be put off by that. Cries of, "this game is cheap!" and "this AI cheats!" are almost guaranteed when you have a game that's actually difficult, so I think a lot of devs instead shoot for the sweet spot where the AI still poses a challenge but isn't too hard to the point where you're constantly loosing.

Consider the anecdote about Sid Meyer designing the first Civ. In battles, a play tester would see that they have a 70% chance of winning, so they'd go for it, loose a lot (say, 1/3 of the times) and get pissed off. 70% to them felt like a pretty high chance, so even when the AI was behaving according to what it blatantly told them, they felt cheated. The compensation, was that Sid made the game say 70%, but actually calculate something like 80%, to ensure people were having the experience they expected.

It gets exponentially more complex in a shooter, so imagine the frustration when the SWAT team you're up against behaves according to real tactics and real experience that you (as a probably not military trained person sitting on your couch) don't have. You'd be shot dead instantly every time.

Again, I'm not arguing that current enemy AI is perfect, but when we talk about progress and improvements, we have to realize that it's not just a straight path that programmers have shied away from for some reason. Making better AI while keeping the game enjoyable is a much subtler art than it might seem.

Well i would not shoot it down that much. It more to it than just timing, the AI is programmed to react to things the player do. This is not deep in itself but it comes off as higher AI function. LIke if you try to use your flask the enemy will often try to be more aggressive. Or if you are to far away the enemy may try to heal itself. The AI does not adapt to these situations if things change which to your point make them exploitable. I feel it just demon/dark souls has a smaller pool of enemies on screen and in one location which allows them to buff up the AI programming vs having monster closest that have a ton of enemies run at you.

I can all but guarantee that Titanfall won't be a game that pushes AI forward. The AI in that game is meant to be canon fodder, they're there to be killed easily by the human players so they can get a titan faster. I'm not saying its a bad thing, I have grown to really like the idea of AI creep in a fps actually, but I don't think they're going to be breaking any ground. If the AI basically wants to die I don't consider it very good.

I've seen a lot of ai staring at walls, running directly into gunfire, or grouping up to be killed easily by titan rockets in the alpha footage....

The titans themselves can be controlled by an AI if you're not in them which is cool but you're still giving it orders, it isn't acting on its own.

That's a very difficult question. If you mean, have we stopped pushing for innovation in that aspect, then for the most part I'd say that it has, yes. Games just don't put any work into making interesting and dynamic and challenging AI anymore. ESPECIALLY the big popular genres. Hell, even games like ArmA don't do much at all with their AI.

However, we have not peaked, we have not hit some sort of boundary because it's harder to get much better. It's actually not. Hell, some of the best AI I've seen was modded AI in ArmA 2. Halo 3 probably has AI 3-4 times better than any CoD game that has come since H3 released. But games have turned into "hide behind a rock and run around like an idiot" for enemy AI. Yet in Halo 3, there were group dynamics that impacted behavior as an encounter progressed (ie parts of that group were removed or added) and complex actions such as maneuvering and using the multitude of tools at their disposal against the player, and a whole lot of it all going on at once. In Halo 2, the brutes were just damage sponges that ran directly at the player. In Halo 3 they were smacking objects out of the way or clambering over obstacles to improve their position on you, they were coordinating with other brutes, they were using equipment against you or for their own benefit. And as you carved through their ranks they changed their behaviors.

And there have been folks working on modifying ArmA's AI to greatly increase the complexity and tactical ability of the AI in the game. AI gained the ability to making bounding maneuvers, suppress and support elements, maneuver to flanks, search an area if players pulled back and reduced their presence, throw smoke to cover an advance, etc.

I actually think more than anything, this is a manifestation of the insistence by the industry that players don't want to be engaged and challenged and pushed back on. So difficulty scales with damages and health values instead of aggression and tactical ability. Ultimately you can give an AI a pretty giant number of actions based on a large number of triggers. That isn't hard, exactly. Developers just don't want to push players to be more aware, more tactically minded. They just want to let them pull the trigger and then pull the other trigger and watch bodies fall.

Someone also made the point about Ai just having to exist in more complex environments and that is certainly another factor in the AI situation currently. AI has to be capable of doing basic things in environments that are much less conducive to those basic things than the environments of 10-15 years ago. Games were less dynamic, smaller, and less detailed. Now there is a lot going on for the AI to cope with just to function on a basic level. It doesn't make things VASTLY more difficult, but certainly provides an extra level of challenge.

That is pretty much you saying that is has stagnated.

Although, I disagree that it's all that hard. We don't have much in the way of even relatively complex AI's. It takes a lot of work but it's not so much that it's hard.

I'm of the theory that truly great A.I. wouldn't be fun at all to play against. It's all about finding the sweet spot for the experience.

However I hope we see improvements this generation developing more complicated behaviors that make the AI seem more interesting than it really is.

And I want to fight a Dragon that behaves like a "real" Dragon would at some point this gen!

P.s.: you could replace the word Dragon by Dinosaur in that last sentence and i'd be even happier, but no one is trying to make a Dinosaur game which saddens me deeply.

I think saying design and creation has hit the rocks is extremely blinkered. Graphics have improved so much compared to AI because for years now hardware sales have provided the money and incentive to create robust graphics APIs and libraries that everyone could take advantage of. When you want to make a game today you don't have to write your engine from square 1, you can license an entire engine, leverage OpenGL, whatever - but that isn't the case with AI.

In fact I think comparing Graphics and AI like that is part of the problem - a good renderer can render pretty much whatever the artists can throw at it, and so a great engine can be used to make all kinds of different games. AI doesn't work like that, the AI in f.e.a.r. is only good at doing one thing, being the AI in f.e.a.r. There is no actual intelligence there, it's just highly domain-specific code for dealing with the things in that game and that game only. That's why compared to graphics libraries, AI libraries are virtually non-existent. In a sense for AI it is a lot closer to starting from scratch with every new game and it's just as hard this time as it was last time and you can't license something from the Unreal guys to do 90% of the work for you.

It's a lot harder to use better AI to sell hardware (so it was always going to take a back seat compared to graphics when it came to where the money got spent) and it's fundamentally a much tougher thing to do than graphics - I don't think it's necessarily true that if AI had been given as much money and effort as graphics that it would be trivial today to implement great AI in all our games, they're just different kinds of problems and you shouldn't expect progress in both fields to be easily comparable.

TL;DR - it's not that everybody thinks making shiny things shinier is all that matters, it's that making shiny things is so much fucking easier than making more 'intelligent' AI for our games and to compound that there's been a historical financial incentive towards making things shiny that hasn't existed for AI in games.

*tells miranda to hide over there away from collector fire*

*miranda stands up onto a crate, gets shot, grunts a lot, doesn't move, dies*

can anyone fix this please?

possible gotg over here, we should make sure, ye know... it's not a complete moron.

anybody?

*wins gotg anyway cos miranda takes her top off in act 3*

I hope that in my lifetime AI gets to the point where I can have my own personal Scarlett Johansson voiced operating system that loves me like no one else does.

At that point there will actually be cool motion and voice controlled games too. I hope I live that long as well. I thought that movie was incredible by the way.

Yeah it has, and until we stop faking it with little technical tricks, I don't think game AI will develop at all. Then again the next step is probably actual intelligence, which is very far away.

In terms of the actual intelligence part of A.I. then yes I think it has stagnated in the sense that it's rarely focused on compared to other aspects of game design (in action games at least, where it could potentially be the most impressive). Dark Souls was mentioned further up, and whilst it would be difficult to describe the enemies as smart, the developers sidestep the issue of 'dumb' A.I. for the most part by having great variety in the enemies. Instead of the player having to react to the thinking of the enemies, it's instead the steady supply of new enemies with different attack patterns and animations which keeps the player on their toes and the enemies interesting, for the most part, as opposed to just fodder (though that changes once dying forces repetitive playing of a certain section).

@cybertification: I don't mean a game like Farcry 3, which is nominally a shooter; also I can't think of a recent first person stealth game that was actually a stealth game (Dishonored maybe? but that's Bioshock AI not stealth AI) like The Last of Us, Metal Gear Solid, or Tenchu.

Is it possible that there's only so much that can be done before the benefits outweigh the cost?

AI isn't as great as it could or should be, but for 95% of gamers, it's perfectly fine and we should all know by now that the minority gamer gets shafted in this day and age. Big companies don't give a fuck and small companies don't have the budget.

I'd probably kill myself if I had to sit at a computer all day fighting with getting an AI script to work and test it out.

Seems to me that no one has much interest in improving it. It's disappointing. Consistent, though, with the increasingly half-assed approach to gameplay mechanics implementation in the past several years. There's no innovation outside of the small budget and indie scene right now.

Sometimes it feels that way, but the AI in Dark Souls is actually pretty smart (most recent observation) so... not always. It's just not often that you see AI you like to call out.

I think it's more about how AI is used, than actual AI level of intelligence getting better or worse. There isnt a lot of creative uses of AI. All video game AI is really going to basically be the same, though.

The secret is that they could make AI better but they don't because they know that if they make it too good it will become self aware and kill us all. Follow the money.

What are you talking about? Every single multiplayer-focused game has actually featured Turing-beating AI for the past 4 years...

*play Twilight Zone music*

FEAR's AI still impresses me to this day, and the AI was pretty much really clever scripted events. You can break it by just rushing through levels and they won't know what to do, but still.

AI is "good enough" for the types of games that the majority of people want to spend money for. There is no motivation to improve it. In fact better AI in, for example FPS games, would probably frustrate and confuse the average consumer. They are more or less expecting a shooting gallery.

I'm no expert but I get the impression most AI in games today isn't really AI - it's all scripted with maybe a simple state machine.

To be fair, the AI companions in ME3 were much better. In ME2, Shepard would almost always be the last person alive in a fight but ME3 they actually do take cover/change covers when being engaged. They also listen to your commands much more accurately when you tell them to take cover at a specific sport. That's why I still really enjoyed ME3, it was the best playing Mass Effect out of the series.

In terms of AI in general I feel that is a tricky issue. You don't want the AI to be too intelligent, else you'll make it seem like the AI are superhuman who can see through walls (for example in an FPS where one guy spots you then everyone within the vicinity know where you're hiding) but then if they are too simple they don't offer any challenge.

AI just isn't a huge focus in games development. Firstly because it's an improvement which is hard to market and Joe Blogs wouldnt fully appreciate, and secondly because it is just plain hard.

There's only so far that a complex binary state machine can take you before it gets overwhelmingly complicated, and alternatives such as genetic algorithms are just unsuitable for games development.

I don't know, half the time, I wonder if it's right or wrong to not see what the AI is doing. Because right now one of the biggest problems is we need set piece moments, so the AI is dumbed down so the camera can get an "AWESOME" shot. Like Bioshock 1, had enemies run away and run off to heal themselves, but then people would totally miss that, or forget to rig the Health dispenser and completely miss out on seeing that explode on them and call them as silly enemies getting themselves killed.

Yes stuff like that isn't that hard to program but can look good. It's essentially just some simple if statements, like 'if player is >25m away and hp<25% then heal'. We had plenty of process power last gen to have decent looking and feeling AI, it's more game design than AI intelligence.

I found the AI in Fire Emblem Awakening to be pretty good at focussing on a target and taking your team down one member at a time. That's the biggest one that springs to mind from a recent game... but really I don't need most AI to be very good. I play online for the competitive challenge and single-player for a different kind of experience. Generally speaking.

AI needs better middleware and probably a different chipset to advance beyond simple scripting. We need to see a revolution similiar to the emergence of separate GPUS for graphics if we ever will see proper AI in games.

No AI hasnt stagnated. AI in shooters? Maby. I think it's because most campaigns are so tightly scripted it doesnt seem like there's much point in making hostiles anything more than gun fodder. Hopefully Destiny can advance things a bit. But it's an MMOG that can supposedly run on last gen hardware so i'm not holding out much hope.

The only games that require a true AI are ones where you're asking a computer to sit in for a human player, so games like Civilization or Mario Party, or, to a lesser extent, in FPS games with bots in a simulated multiplayer setting. But when it comes to asymmetrical single-player experiences where the role of the human player is different from the role of computer-controlled antagonists (your character in Dark Souls and those poor monsters you're supposed to kill are playing very different games), then all you need to do is make the enemies act interestingly, not necessarily intelligently. Since a majority of the gaming market share is wrapped up in these asymmetrical games, I don't think AI will ever be front and center. As a strategy gamer, it's something that I care a lot about.

Here's a short 2012 piece on AI from Stardock, who claim that their Galactic Civilizations was the first commercial game release to feature multi-threaded AI.

You're probably right but it would seem to be a good vehicle to push AI forward. It should have an option with an AI slider from cannon fodder up to a formidable challenge. There's so much real world data to pull from when it comes to online fps, I can't believe some programmer out there can't use that and build a really competent AI.

The big problems with AI is that so many of our games are confrontational, thus improving AI is just saying the enemies in a game will kill you faster. But killing you faster is not fun. Right now game developers can make AI that will beat you fair and square every time. AI that will run around a race track better than you could, AI that will see the second you move on a shooter map and kill you, etc.

What has to happen is for more games to be experiential, where that game play is about experiencing the world and where the AI is made to heighten the experience. I think in RPGs you could do a lot with trying to model interactions between NPCs and your character. You could have a game where some of the dialagie is part of a pre-planned tree yet all teh NPCs could talk about anything. Imagine meeting an NPC on at a farm in Skyrim and a dragon flys over and the NPC in the middle of their discussion to you says, "Holy crap did you see that dragon! I know I was talking to you about my missing onions...but fuck it I don't care about onions when a dragon is flying around my house." Or how about NPCs that at teh start of the game might give you a quest, but later on when you are the MOST IMPORTANT knight/mage in teh kingdom might just bow down to you or ask you to kiss their baby.

The problem with AI now is our games are too simplistic and too narrow, so to improve the AI on such games is a waste. Better AI will only come for games that have very sophisticated worlds and very sophisticated interactions within those worlds. If you want sophisticated AI in a shooter than there needs to be more possible in the game than walking, shooting, and ducking. You could put amazing AI into a shooter where you can grab someone tie them up in a broom closest torture that person for 30 minutes while they scream about their kids and wife before you slit their throat just after then tell you where the secret plans are hidden...but fuck it that would probably give everyone bad dreams.