When do you think AI will dramatically improve?

>>383866493
Never.

>>383866493
since it seems to be the only thing that stagnated across all platforms im going to say never or AI will require an internet connection and run off massive neural networks that are taught to play and developers rent it for players so again never

No, it won't.

Holodeck games or something similar at which point the inadequacy of in-game agents becomes truly jarring for everyone and market has no choice but to adapt? An easily trained general-purpose (but not general intelligence) neural network such as DeepMind becomes cheaply licensable? Artificial general intelligence (either as a fellow person playing the game or superintelligence churning out top-quality personalized experiences as easily as a human coder writes a "Hello World" program)?

As far as anything resembling the current mainstream market is concerned, AI is unlikely to ever become a major selling point. AI isn't marketable because you can't advertise it with screenshots and gameplay footage might as well showcase fully scripted sequences. Moreover, a lot of people don't WANT good AI opponents because, even when the opponents play by the same rules as the player, any degree of competency tends to be perceived as "cheating", and people like asymmetric balance anyhow (beating tough "boss" enemies an order of magnitude more powerful than you are or dozens of their minions, feats that can only be accomplished if you have some sort of edge, that being exploiting their dumbness).

I just got the greatest idea for the ultimate co-op game with super advanced AI. release two versions of a game, let's say it's left 4 dead 3.

version 1: released to the NA, UK, AUS and other "first world" regions. in this version it's exactly what you'd expect, you play as the survivors, or in vs the special infected.
version 2: released to the "third world" regions BR, SEA, etc. under a different title and F2P. in this version the players are the regular infected, and occasionally the specials.

both versions interact with each other via online connection, but neither region truly realizes this since there's no communication between players cross-version.

>>383868665
>what if every enemy had 300 ping
great plan, I'm sure it will do so well

>>383866493
it won't ever be good

>>383868919
just make it so the netcode favors the version 1 clients.

>>383869084
You have no idea how latency works over distance. Go away

games need AI that are predictable, not smart
smart AI fucks up the gameplay by ruining your plans and being generally unpredictable. it sounds like it would be fun, but it really isn't

>>383866493
Never because everyone just cares about graphics

File: 1497702860756.jpg (10KB, 400x400px) Image search: [Google]

10KB, 400x400px

>>383866493
Games are becoming an increasingly multiplayer affairs, and AI is cpu intensive. Guess what is getting cut first.

>with all this extra power we can finally have the AI we want
>still behaves like shit

It's programmed by talentless hacks to appeal to casuals

Even if someone could make good AI, the publisher would tell them to fuck off and just get the product out asap

>>383866493
none of that shit even matters since the player never sees it. what's more important are interesting scenarios.

>>383866493
>wanting ai to improve
>wanting the human race to be wiped out by machines

>>383866493
never, it's always more economical to only create a rudimentary micro-level AI and script everything else manually

>>383869876
Or if you're Naughty Dog and you're making The Last of Us, you don't bother scripting and call it a day at the rudimentary AI.

What game has the BEST overall AI?

I find it funny when stupid people think games need better AI.

>>383866493

Depends on what you mean by improve. It's easy to create smart AI. Just make something that knows your every move and give it instructions to counter every move you make and to think ahead of you. The real challenge is making AI more "human-like" Imperfect, but realistic on what a human would do. That kinda shit is hard to make.

It already did. In the mid-late 2000s.

Guess what? It's memory intensive. At the close of every AAA game cycle you begin picking and choosing what to cut and what to keep. AI is almost always dumbed down. I've seen it happen on several well known projects and known many AI engineers who quit the business in frustration.

The reason is that no one is going to accept a game that looks one to two generations behind graphically no matter how good the AI is. Or at least, that's how publishers think people will react. And that is all that matters really.

As an example, on BioShock 2 we wanted to have the Big Sister (before it became multiple entities and was just one character) legitimately track you throughout the game, SA-X style. We had three levels rigged up to our engineer's stalking system and it was fucking amazing, battles came out of nowhere and it was fun gameplay. But it was pretty memory intensive, required level architects to do a lot more work in an already old engine, and ultimately production felt we could get 80% of the same bang for our buck if we just scripted each battle and gave the Big Sister some basic attacks.

>>383866493
Do you think that "intelligence" is the same thing as fitting a curve to data? If so, artificial intelligence is here. If not, artificial intelligence will never happen.

File: IMG_7199.jpg (382KB, 2048x1896px) Image search: [Google]

382KB, 2048x1896px

>>383868919
no because they would not play like mindless zombies just mindless third worlders

>>383869084
Won't work, retard, version 1 has NA, UK and AU. Latency will definitely remain high due to their distance.

>>383870392
>consoles

>>383870392
And it was shittier than the first one, good job.

>>383869273
isn't that what they did with stalker at first? and it went to some weird places.

>>383870392
what are you doing that makes it use so much memory? even complicated AI maps only use as much data as a few textures would

>>383866493
Noticeably clever AI just can't be received well.
Take any time you've played even a barebones form of a game like online chess where the hardest difficulty level computer destroys you in a matter of seconds.

Beating it never feels rewarding, it feels like a struggle and losing never feels like a lesson because you can never be sure in compartmentalized chunks, something only possible when you dumb down and isolate the recognition, reaction and response of the enemies, where exactly you fucked up.

Chances are you have more fun when the computer intentionally gimps itself, when it's not reacting to your every move and mistake with accuracy and intention and gives you what seems like a "fair" chance.

So as another anon said what really sells is AI that looks smart, not is smart. AI that is programmed to behave in ways that you the player can recognize as demonstrations of intelligence, even if said displays don't necessarily entail intelligent decisions, and the most intelligent decision a company can make is appeal to the audience not upset it.

tl;dr You think you do, but you don't.

>>383870136
why?

AI that is smart enough to hide behind cover while being shot at is smart enough. What makes bad AI is not reacting to your surroundings or current situation and adapting. If it does that 60% of the time like most AI in games I've played, I have nothing to complain about. It sounds like to me people just want harder and unfair AI.

>>383866493
The problem with advanced AI in videogames is that we don't need advanced AI in videogames, we just need AI that seems clever and convincing to the player, because AI is only there to support the gameplay.
It can never "transcend" gameplay and must always "think" within the confines of easily perceivable game boundaries, otherwise it risks ruining the player's gameplay experience. To answer the question: AI will improve when games will become more complex and require more complex AI.

>>383869273
It would have application in fighting games and for bots in shooters. In those you want your opponent to actually fight against you and be a real threat because it's a substitute for a player.

Also, smarter AI can be used on a lot more than just enemies, and can be used on enemies conservatively for basic shit like making sure they don't bug out doing something as simple as trying to navigate the environment.

>>383866493
>American college

Until machine learning really takes off you're going to be stuck with similar levels of AI for a long time.

Also what >>383871154 said. People tend to react negatively to AI that actually does what's most effective in every scenario because it feels unfair, and it kind of is. Which just makes AI programing all that much harder because you basically have to simulate human stupidity to an extent and that's difficult to do well.

>>383871429
you can make a bot in action games unbeatable just by giving them perfect reflexes

>>383871314
>harder and unfair AI.
more like AI that can't be exploited

File: 1480563589344.png (211KB, 496x379px) Image search: [Google]

211KB, 496x379px

>>383869369
Multiplayer?

>>383871586
That's why I said "smart" with quotation marks. Like this person said
>>383871536
one that can mimic human error as well as unpredictability. Ideally without frame 1 reaction time.

>>383871723
I don't know about fighting games but shooters have had reasonable bots for decades that even learn

File: shia_mug.jpg (220KB, 900x494px) Image search: [Google]

220KB, 900x494px

Battleifled 1 has the shitiest AI i ever seen
even cod has better ai

>>383871429
a competent ai would never lose in a fps or fighting game.

those are both games based on reaction times and ais have a reaction time of 0

>>383868919
>>383869265
>>383870675

Just have 3 servers.

Americans play as Mexicans and BR's. Brits play against middle east and north africa (low ping too since they can play from Germany) and Aussies can play against SEA nations and NZ.

GOAP was state of the art when it was introduced in FEAR, but that game also used a few other tricks

Good AI makes the player feel smart, it's more than just how to be obstacle and many games have forgotten this

AI's role is not to outsmart the player, but the present interesting gameplay.

>>383871185
It depends on the game, but let's look at Chess.

Now, contemporary Chess engines don't actually calculate every possible move - in fact, they prune the lines they search VERY aggressively (single-CPU consumer hardware is only now beginning to catch up to supercomputer hardware Deep Blue had when it beat Kasparov, but engines like Stockfish and Komodo running on a regular home computer would be like 500 ELO points ahead in strength), but because Chess is a turn-based game with perfect information, it's easy to just search through all legal moves (the kind of difficulty that might appear as batchelor's degree computer science exercise). Doing so would enable them to look "only" ten or so half-moves ahead (as opposed to 30 something you might see with modern engines) but that kind of tactical play would be good enough to handily beat any amateur. However, that is completely unlike how a human would play. Even grandmasters may lose to mate-in-two or mate-in-one (such as reigning world champion Vladimir Kramnik against Deep Fritz) because humans don't look into all positions, while such a computer would have exactly zero strategical and positional sense (neither do contemporary engines, but they calculate their chosen lines so deeply it appears as though they did). It simply wouldn't lose material in short sequences and would see every possible mating line within 5 turns.

Conversely, you could look at FPS games. It's easy to make a computer agent that, in every processing cycle of the game, checks if it has line of sight to a player, and if it does, presses the trigger. In a situation in which a human player would see one pixel of the approaching enemy if they decided to turn 180 degrees the exact frame in which this one pixel becomes visible, the bot would already have landed a headshot.

>>383871185
Because machines don't think like humans.

>>383870987
2 was better than 1 in every respect you fucking plen.

>>383868665
The developers of W40k Eternal Crusade presented a similar system during early development stages, where people who bought the game would take the role of SM, Eldar, or Chaos and the unwashed masses, who picket it up free to play would be forced to play greenskins.

>>383873145
Now, not nearly all games are like that. For example, Go is also a turn-based perfect information game like Chess, but just two years ago it was thought an AI capable of beating high-level human pros was a decade or more away, and while AlphaGo now reigns supreme over humans, it wasn't through "easy" (to program) approach of calculating all moves but neural network playing millions of games against itself and gaining a similar sort of "understanding" of the game as humans have.

Many other games pose even greater difficulties. For example, one might naively think that a computer microing at 100k APM should trash humans in StarCraft but that actually isn't the case. Brood War AIs have been a point of research for a while now and having looked at AI tournaments, even I would beat most of the competitors and a human pro would do the same without breaking a sweat. It's easy to write a script for microing mutalisks in a very specific scenario (say, fighting against marines), but it's difficult to make an AI that manages all scenarios (such as a mix of marines, goliaths and science vessels) and even more difficult to get to the point where the superior micro becomes a factor.

>>383872864
AI's role is not to outsmart the player, but to look like it almost did so the player can feel smart and think he beat a difficult boss.

>>383873220
Machines don't think at all; you might as well ask a pocket watch what it thinks about tactical gameplay.

>>383873539
hmm

>>383874083
No, they don't. You are correct about that. But similar to the chinese room, there is a hypothetical point at which the emulation of thought is becoming so close to the real thing, that it will be indistinguishable for almost all relevant purposes.

>>383873145
ah.
>>383873220
not much of an answer

>>383869273
This. Star Rangers 2 has an AI that knows how to game the system, and thats just plain anniying.

>>383874430
That will be the point at which we can safely call psychology a "solved problem," which will also be the point at which we can call philosophy a solved problem. This current hype of going to town on regression analysis is not going to get us anywhere near what you're talking about.

>>383870081

>What game has the BEST overall AI?

If you read this:

http://alumni.media.mit.edu/~jorkin/gdc2006_orkin_jeff_fear.pdf

You'll get a nice glimpse into how the approached AI for FEAR and how they used certain tricks for the AI crapping itself in loops or stalls in the AI chain.

>>383867386
Why though? You can literally make a neural network that learns how to play games on your PC, it's not that hardware taxing. Devs are just fucking lazy as fuck, the AI didn't stagnate, it outright regressed from older games.

>>383870081
FEAR
STALKER

AI will never improve so long as the market is geared toward multiplayer

File: hlep.png (301KB, 573x433px) Image search: [Google]

301KB, 573x433px

>>383876935

>tfw I used to like multiplayer
>used to fucking LOVE multiplayer
>now it's not only overdone, but homogenised to shit and catered to the lowest common denominator

Remember when you could launch practically any decent online game or mod and find at least a server or two of people playing? Now it's about 5-10 games, discounting MMOs (which are all identical anyway).

>>383870392
PCs now have 4-8x more memory than PCs from 10 years ago, and AI is still regressing rather than progressing, so I think that's not it anon.

>>383872772
Why isn't goap used more?

>>383877330
>Now it's about 5-10 games
You can launch practically anything that isn't flavor of the month from 15 years ago and find people playing it. The top 5-10 games on Steam alone all have over 10,000 players, the number only titans like Counter-Strike could surpass back then.

Precisely nothing had changed over years, including games that are actually popular appealing to the lowest common denominator, people complaining about it, and allegedly good games being dead because of it.

>>383879745

I should add that those people I found years' ago weren't off in bumfuck Europe with 300 ping

>>383879729
Might be proprietary code, also it might only be good at that one specific thing.

The idea that we have taken metals and elements from the earth and synthesized them and turned it into machines that turn electricity into moving pictures that can be controlled by a human being and which in turn contain characters which can dynamically react to your actions is like fucking magic to me. I cant even comprehend it

>>383876793
Stalker AI is complete trash.

Well now that graphics have pretty much peaked with realism I think the industry might shift *some* focus on AI. It's probably going to be less 'AI' rather than ' NPCs now have more ways to interact with and have more reactions' level TECHNOLOGY like zelda(terrible example I know)

if the a.i is actually good people always accuse it of cheating

you cant win against an a.i if its good

>>383880096
GOAP isn't proprietary. Jeff Orkin created it for FEAR, but the architecture has been used in more games later on (STALKER, Deus Ex etc).

>>383880487
You are retarded.

>>383870081
The gooks in Crysis 1 were pretty smart

>>383880602
>peaked with realism
They haven't though.

>>383880602
Graphics haven't improved in over 5 years

>>383881043
Oh no, looks like I hurt somebody's feelings! It's true, though. They don't work in groups and they completely shit themselves when they have to fight more than one enemy (so every time you have allies, basically). They're really bad with switching targets, let's say an enemy NPC is shooting a friendly one who's at like 100 meters away, you can just walk up to that enemy and start spraying him and he often just won't switch targets. Fucking idiotic.

>>383881148
They can be pretty impressive when they aren't glitching the fuck out. A lot of their reflexes are artificially lowered even on Delta, though.

>>383881295
that's what I'm saying, graphic improvement, while still improving bit by bit, has been a total slog. devs will eventually have to search for other things to 'innovate'
>>383881263
getting mixed messages here

>>383881263
Elcideon engine never ever friend.

>>383874860
Thanks for posting this, giving it a read.

>>383876132
>You can literally make a neural network that learns how to play games on your PC
Yeah, but it needs a shitload of time on big clusters to do that, and you'll be lucky if it does learn something.

And you wouldn't even know how exactly the system will behave if you train it with neural networks. Maybe a palyer making a circle movement will freeze it forever? You just don't know. Neural networks are good at a certain kind of tasks, but AI isn't one of them.

>>383881935
Sorry, I understood you poorly.
But I still think we can improve, mainly on lighting