So somebody got an AI to play a videogame using only screen data and nothing else

Because it would outclass you entirely. It would just aim at you with 100% precision from the distance you won't even see it at. That's why the best AI systems are the ones that manage to seem human, like FEAR soldiers and UT bots who rely on communication to corner the player while simulating actual person's aiming and reaction.

It's not magic

Dev here

This is actually pretty big. Will certainly lead to some big advances.

Most AI in games works like a cheater in CounterStrike. They don't see the picture you get on screen, they basically have wallhacks, aimhacks, etc, and are then dumbed down to fake an approximation of human skill. It's why CSGO still has bots that will 360 noscope half their own team when flashed - their 'dumbing down' system fails.

BIG gains here in terms of tactics AI can deploy, because you can train them, the same way you can pre-compute lighting and occlusion, rather than have to script behaviors on a per-room basis.

Also massive gains for anything that uses camo - instead of an AI knowing where a person is and adding abstract modifiers (distance, color, etc) that often don't make sense in practice, it can be trained with cnn's to recognize things.

Hopefully this brings on a sort of power creep as AI can push development of tactics and gameplay to new levels.

So do you think this will actually let a game be more fun? Obviously the AI will need dumbed down as you say, but could it lead to less "that's fucking bullshit" moments?

the difference here is that the AI is controlling a player, in most FPS games your enemies are not remotely similar to your player character in ability or capability.

I honestly doubt it. Think about it, you need to actually render an unique full frame (including lighting and texturing if you want camo to work) for each and every single bot you have running, plus the one for the player itself. Basically, this rules out any offline application this might ever have since no home computer would be able to handle the load if the game is even remotely graphically intensive.

I got you familia.

Hopefully yes - bullshit situations now are 90% cases of when the dumbing down method fails. One bonus that a lot of people skip over is that players are a lot more accepting of losing against a human (or what they think is a human) than tying or slightly edging out a strange, unpredictable bot.

Aren't they? In ARMA, they're the same, as is CS, and it could probably be said for the majority of FPS games as well. Obviously they're all MP games, so players need to be on opposite teams o other players, but how many recent FPS games wouldn't benefit from better, more logical AI behavior?

To a degree, yes, but this is something that has had to be improved recently anyway, mostly due to VR and actually considering the capability of a human eye and how we deal with an image. Aspects like FoV, pixel density, etc are now coming into focus and people are working hard to try and get something that solves what is an Eye problem, not just a human problem. Additionally, baby steps can easily be made by managing rendering resources. I'm not saying you should expect Middle Earth level battles with this, but a 8v8 isn't out of the question on modern highend hardware, especially if they can optimize it.

Also, cheesing remains the gamedev's best friend; for radar in a flight sim, you don't need albedo, but between IR/Thermal, Reflectance, absorbtion, velocity, etc you could almost get away with one RT (rgba) pass per plane to cover >90% of radar methods used by militaries right now.

youtube.com/watch?time_continue=2&v=9qWHM8DNdr8

this is an AI that plays perfectly (fox) vs a top player. He literally can't touch it once.

Wow they out a health pack in its path finding every 5 seconds so it would not die from running head first into every enemy woooooow

Boston dynamics needs to watch the fuck out

Any data scientist/engineers itt?

works """"fine""" for arena shooters, but the most recent wolfenstein and doom would not really work out the same. Enemies arent controlled the same way as the player this is true for id say 90% of all FPS.

The question is how the bot was programmed.

>manly tears plays multiplayer.

>this is something that has had to be improved recently anyway, mostly due to VR and actually considering the capability of a human eye and how we deal with an image. Aspects like FoV, pixel density, etc are now coming into focus and people are working hard to try and get something that solves what is an Eye problem, not just a human problem.

I don't see how this is related to performance. The only tenuous link is how VR reuses certain things to make the second frame cheaper to render, but that doesn't necessarily apply here since bots could be in entirely different areas of the map without any overlapping elements in their viewports to reuse.

Even then yeah, provided a small enough bot count, a graphically simplistic enough game, a powerful enough computer, a graphic engine optimized for rendering multiple viewports simultaneously, enough corner cutting (running the bots viewport at drastically lower FPS, possibly even resolution, lower color count, etc.), and an advanced (and fast) enough optical image recognition algorithm you might just be able to successfully improve on a minor feature that's outside the primary scope of any multiplayer FPS.

To be honest this is a solution in search of a problem. As a curio it's pretty cool though.

I don't know what you're talking about.

The decision-making process, that tells the AI to "go there / target this / etc" or the actual movement that then takes that command and tries to execute it (check navmesh, rotate, jump, run).

Because in terms of actualy moving a bot around, it still needs to walk, jump, etc, and even if some particular game uses a different method (shout out to some fucked up examples, like just lerping), that's irrelevant.

What we're trying to improve here is the decision making process by better aligning the inputs to something that makes sense. And this way, while still in its infancy, is a fantastic first step.

I wonder if through the process of evolution the bot actually learned that the smartest choice was to choose Fox, or if it was just designed to choose him.

>that part at 1:40 where it was unsure if it wanted to kill or pick up more ammo.

It was probably designed to play Fox, I doubt the programmer went through the trouble of making it able to play every single character perfectly only to sit back and let it decide what it wants to play.

Just let the dam deep learning bot play your game for a few months in advance, it will figure its shit out on its own I'm sure.

Sure most people have seen this, but this is an interesting video that gets the concepts of neural networks and genetic programming across in a really simple way.
youtube.com/watch?v=qv6UVOQ0F44

it was chosen to play fox because he has the most broken attack in the game (shine) which comes at frame 1. There's nothing faster than frame 1

I'm not just talking about the hard technical side, but also things where pixel density would matter (center of vision) that would allow these extra bot-vision renderers to run at higher quality where needed and squeeze out some extra bang for your processing buck. There were some old projection matrix buggery that sat abandoned for ages before VR came back to the light, and like in VR it's all about maximizing the quality for where the eye can see, except it's for a bot's eye.

This likely won't come out and immediately blow current bots out of the water in terms of max-level skill, current bots are too cheaty for that anyway. What we should get is a bot that has more human behavior, due to seeing through what is similar to a human's eyes.

As for pattern recognition, lube up assholes, because our good friend FXAA (and TXAA, and SMAA, and all that) and its development has meant a lot of work in trying to analyze images fast, even on consoles. The same algorithms that try to find jaggies that everyone mocks might just be the start of ultra-fast pattern recognition.

That's one method, and probably the most 'safe' one considering you can see and over-ride things before release.

Microsoft tried to do something like this where they uploaded races from Forza on the xbox onto their cloud for analysis, and I think a Halo game (odst maybe?) and an EA game (titanfall or earlier I think) also did something like this.

Unfortunately, if you feed your AI replays from the tards on Live, you just get retarded AI.

Microsoft, having learned from this, then made that twitter bot a few months ago that turned into a nazi.

I still miss Tay :(

>To be honest this is a solution in search of a problem. As a curio it's pretty cool though.

Actually, I think the problem it could help with is less actual enemy AI, and more that you can do free playtesting.

Why hire playtesters when you can get an AI to run through everything a gorillion times?

Which is why it's kind of meaningless. I want an AI to be able to learn how to play best, not just be told how to do it.

You just made me consider, if you're making an AI that can see like a human, would it also need to understand the concept of focusing?

Surely seeing and comprehending the entire screen at once is not actually something a human can do.

because programming casul bots to be a dummy for casul players kills the programmer

For people who aren't casul, you're trying to take a system that was built to your level and make it as good as a 5 year old or a 85 year old. There's different problems to solve there.

Otherwise, yes, sometimes we already use bots or other analogues to try and run through and break shit. This stuff is a bit more advanced than that though.

Waifus

sort of.

We can take some things for granted - since the screen doesn't track 1:1 with a players eyes, we can turn DoF and such off for the player because it can blur things that people want to see but aren't aiming directly at (otherwise your crosshair would by flying all over the place as your eyes move around the image)

Likewise, and this comes from the VR stuff, we can assume that there's a good chance that what they DO want to focus on is probably near the middle of the screen, so we can save a decent chunk of power if we can increase the pixel density near the middle (for better analysis) while still having some lower quality peripheral vision.

With neural networks making massive strides in AI we'll soon have AI waifus to love us.

Their voices are already improving: deepmind.com/blog/wavenet-generative-model-raw-audio/

youtube.com/watch?v=JjRyHgF8hx8

Damn that sounds pretty good

The babbling ones are creepy as shit too

>Breathing and mouth noises

storage.googleapis.com/deepmind-media/pixie/knowing-what-to-say/first-list/speaker-4.wav

>those piano things

kek, its like mozart having a stroke

storage.googleapis.com/deepmind-media/pixie/making-music/sample_3.wav
>this will never be a real song

storage.googleapis.com/deepmind-media/pixie/knowing-what-to-say/first-list/speaker-5.wav

Number 5 sounds like a total qt3.14

>We show that the proposed architecture substantially outperforms built-in AI agents of the game as well as humans in deathmatch scenarios.
it's already unbeatable. But you want that in your war machines, not games.
omicsgroup.org/journals/genetic-fuzzy-based-artificial-intelligence-for-unmanned-combat-aerialvehicle-control-in-simulated-air-combat-missions-2167-0374-1000144.pdf

It's been in war machines for years

Sounds like what I imagine all slav women sound like

it has no tactical awareness, after someone leaves its sight it forget it exists
if it wasn't playing doom with some really bad players it would suck dick

>Microsoft, having learned from this, then made that twitter bot a few months ago that turned into a nazi.
RIP. Taken too soon.

Nobody said it was perfect

This isn't the original version of the pasta. This is some shitty roleplay faggotry where the guy decided to act it out on Cred Forums.

Get your reading glasses on and hear the tale of how an 80's AI developed a completely novel strategy to curbstomp a bunch of nerds so hard they had to change the rules of the game.
aliciapatterson.org/stories/eurisko-computer-mind-its-own

>"They changed the rules significantly and didn't announce the final new set of rules until a week or so before the next tournament," Lenat said. "The first year that would have not been enough time for me to run the program to converge on a winning fleet design." But Eurisko had learned heuristics that were general and powerful enough that they could be applied to new versions of the game.
"We won again and they were very unhappy and they basically asked us not to compete again. They said that if we entered and won in 1983 they would discontinue the tournaments. And I had no desire to see that happen." So Eurisko retired undefeated

Holy shit, this is hilarious

In that scenario it's kinda pointless. If you already know what the absolute best character would be (in the hands of an infinitely skilled player) why bother making it learn?

It might sound sexy and revolutionary, but teaching an AI to learn how to play is basically down to statistics. Tell it to monitor a number of variables, perform an action, and record if the result was favorable or not. Eventually it will have accumulated enough data to be able to pick the action to perform in any given situation that is most likely to produce a favorable outcome.

>"To counter opponents using the same strategy, Eurisko designed another ship equipped with sophisticated guidance computer and a giant accelerator weapon."
>"Its only purpose was killing enemy lifeboats."

This sounds like something that would be at home in EVE online.

Isn't EVE a spiritual successor to Traveller anyways? Or am I mistaken?

I wouldn't be surprised if some AI of this caliber came about. Anyone that's good at EVE is either a genius or rich as fuck.

>Lenat called his new program AM

I didn't think anyone would read further than the conclusion of the Traveller tournaments.

I usually finish articles I deem interesting completely. Thanks for the fun read, user.

>"the cell can be realized most efficiently on the surface of a Mobius strip.
very interesting piece, I might have heard of the tournament story before, but with more knowledge of deep learning now it's interesting how it came to be

>implying something programmed to love you unconditionally really loves you

Why not?
I did.

this

I don't mind if it pretends

Yo that's the best episode of the series

I bet a lot of people wouldn't mind if it pretends it hates them

But that's the thing.
It wouldn't have to be "programmed" to love you.

Assuming we're talking about a stage where this thing can develop a personality, or at least likes and dislikes, based on the data it's inputted you can just run enough copies until you get one that likes you.
It's love for you would have been developed as randomly and naturally, as if it was a real person. The difference is, assuming you get 3.5 billion copies of the program, it's easier to weed though all the worthless ones and pick the one that likes you, unlike in the real world where that's not really a possibility.

this only proves how OP the super shotgun always was

If it were sophisticated enough that it could develop a love for you through random occurrence, it's more or less an actual girl. You even said yourself that you might have to look through over a billion variants until you found one to love you. At that point, you might as well try to find a human who loves you so you can have kids if you want.

I wonder if you could start up this bot and play it within a tournament.

>Even my AI waifu thinks I'm a useless unlikeable asshole

THE FUTURE IS GREAT

Yeah that was my point. Your computer waifu would more or less be an actual girl.
But my point also was that it should be significantly easier to search through a billion copies of an AI program for one that developed a love for you, than a billion living human beings.

And then what if there isn't even anyone on Earth who could/does love you?

This.
You don't want them to read your inputs as you put them in and react off of that, then shit would suck.

love isn't magical senpai, it doesn't work like that

Imagine a horror game where you're being chased by an AI like this.

That was fun to watch. It would be nice if we trained board-tans with content scraped from each respective board

How does it work user?

How would it develop a love for you? You don't just run some program and walk away and come back to find out whether or not it loves you. Even if that were the case, you could still argue it's more or less programmed to love you. You'd have to interact with it, talk with it, go on dates with it. How is it significantly easier to search through a billion copies of that?

I think there literally was such a game, something like Hello Neighbor

>It would be nice if we trained board-tans with content scraped from each respective board

I think you're on to something, user.

I dunno, you put in data about yourself or something.
Then it's just a matter of narrowing down to a small amount that find you attractive from what you've given it. And then a smaller amount that you find attractive.
And then hopefully you get along with it.

In a racing game you could set a threshold of only people who won in x time get sent to the ai

Frame 0 is faster

If it's as simple as putting data into it, then like I said, it's more or less programmed to love you. The algorithm that would dictate love would be programmed to find your best qualities. No robot waifu will ever love you, user. ;_;