Why do companies not use Neural Networks to create better and more challenging AI opponents?

Just make and sell texting waifu light switches

>he thinks aimbot is enough
It has been proven that flank, strategizing and controlling the map is too overwhelming for an aimbot to win purely through aim.

>I'm an ML engineer user.
>not know how expensive is to code it compared to some scripted shit

But it is better!

Money is a problem for the eggheads in accounting to deal with, just give me lots of data and a few GPUs.

AI running at full capacity would outplay any human, that's why it's always limited. Imagine being fully aware and being able to fully macro and micro every unit individually of a 80 food army in WC3.

I don't know what you had in mind for "scripted" but (decent) game AI uses decision trees and state machines and npc behavior is mostly dynamic. ML is as much of a sloppy brute force answer to the problem of game AI as it is for most other things it's misapplied to nowadays. I know there've been starcraft bots made using it, but for most games and genres it's probably a really dumb approach with pointless overhead for the same effect.

Machine learning would make the game too hard. AI has already beaten the best chess has to offer decades ago, what chance do you think typical players have against today's AIs?

Theano or tensorflow

>2. You don't really need machine learning to implement great and engaging enemy behaviour.
This. In fact many games probably do already machine learning to generate AI behavior.
But optimized AI does not necessarily make for a fun game, especially if your audience is even somewhat casual.

Where is this girl from?

RL is mostly a meme, but it might be a good idea for some games unironically. A vanilla PPO agent with a decent heuristic reward would probably work much better than the default Stellaris AI.

Not all games benefit from micro and strategic play can be very difficult even if you're going all out desu. A decent human will win without handicaps in many strategy games

The boy you have a crush on is my boyfriend, Nagatoro!

Attached: 1558442324397.png (700x1029, 250K)

>Is your character a YouTuber?

Neural Network is not what you think it is and it wouldn't make much sense in the context of video game.
Also programming AI in video game is pretty much setting some believable and realistic-within-game-rules handicap for opponents, because nothing stops AI opponents from knowing exactly where you are, and how to win over you.

putting together a CPU opponent that could absolutely btfo every person on the planet is not a difficult task nor does it make for fun gameplay

It could be balanced by giving the player advantages over the AI rather than dumbing it down

Seems like its from that manga about a girl who bullies a senpai virgin but deep down loves him.
Please don't bully me, Nagatoro is the name

Games are unironically much harder for computers than Chess because there are so many degrees of freedom and possible states. At least some games. Even openAI's successful dota bots are limited and can't compete with the full character pool. And that is with truly massive investment in training

youtu.be/9jrO58mg-Qg

They've tried, there is a documentary on this, even by giving the AI voice calls over what they were gonna be doing, if they were reloading their strategies etc.. the players would still get absolutely BTFO.

Attached: 1558384265093.png (700x1029, 459K)

It will completely ruin any sense of immersion, for one.
Machine learning is all about finding optimal results, whether in speed or accuracy. The method to achieve such results is irrelevant. That's why many AI solutions reach impractical or unintentional results. This paper has dozens of examples, such as a tic tac toe AI that learned that the most optimal solution was to make a move out of bounds to cause its opponents to attempt to allocate more memory than they could, thus crashing the application and losing the game: arxiv.org/pdf/1803.03453.pdf
In short, the solutions often feel unnatural, and there's a lot of work involved if you want to set the proper constraints to get a desirable solution. This isn't ideal for video games unless the goal is to win no matter what, which seems really fucking boring from the player's perspective.

What's it called? Is it on youtube?

Look at the recent dota2 openAI matches.

Im sorry man i don't remember the source just the contents, they used a shooter for the AI argument but i dont remember which game, it is on youtube but i fail to find it.

Because it needs players to "evolve"
Those players won't play the game until it's 70% done
A game 70% done doesn't have time to dick around with the latest tech buzzword
Especially for a feature that only tryhards that play on "realistic" difficulty play
That won't even get praised because a good enemy and combat design gives visual and audio cues, giving them patterns on purpose so you can win, a faggot that changes his attacks every time will be called "bullshit", not "smart"

Basically it's easier to have good enemy design than "DUDE MACHINES CAN ASSPULL BETTER DESIGN THAN HUMANS, TRUST ME JUST RUN THIS 1000000 TIMES UNTIL IT'S GOOD"

Attached: 1556070048709.jpg (1158x816, 146K)

Derpinaaaa

>I'm an brainlet user

Why would you trade a well understood algorithm that works 100% of the time, does not require an expensive "training procedure" and requires very little processing power and can run on a cpu for a algorithm that sometimes doesnt work for some unknown reason, needs to be trained on massive amounts of data on a set of GPUs and requires megabytes of memory and shitloads of processor time if it can even run on the cpu.

You know that if something renders in 60FPS then you have under 17ms to render a frame. That frame includes enemy AI, game physics updates, lighting, shaders and a whole host of other shit. Do me a favor a tell me how long it takes you to run your resnet 50 on a single fucking 240x240px

This is why noone respects ML people. There answer to everything is "use a neural network lul". In conclusion go fucking kill yourself "engineer"

If every npc was actually smart, you wouldn't be able to finish any game.

Two reasons:
Its not easy to train an AI to act the way you want, at which point you might as well script it.
And because AI that is too good is also not fun.

Nice

>using resnet and image processing for an ingame AI
>when you could just feed the fucking game state in directly
>calling the advanced AI every frame
>instead of just using it to make high-level, strategic decisions
shiggy diggy

>the absolute state of “machine learning” experts
can’t wait for your stupid meme field to crash

Assuming a fantasy land where AI exists and is applied in games to every NPC, if you kill one would that be murder?

Attached: 523346568657.gif (627x502, 941K)

I hate how every fucking field is trying to get on the machine learning train. This shits been around since the 90s. The fad died and it’ll die again

It'd be closer to squishing a fly

he's probably thinking of atari AI, which didn't have access to internal state and had to read frame data

no one will give a shit about computer's rights until we mistakenly teach them how to protest at the UN

From what I've seen most humans win because they are aware of strategies the AI is not. Once they are taught the human strats they dominate.

This. Just because they saw an epic video about AI they think we can learn and act like humans. Preposterous.

ML is peaking again because it's got real world applications now and the hardware can actually run it now
However now they're looking for problems with the ML as a solution

>AI has already beaten the best chess has to offer decades ago

That wasn't an AI. AIs in chess have only been used recently.

>machine learning
more like "Machine repeating things until it gets one thing right"
it's a fad, it doesn't learn anything, doesn't learn to understand concepts and it can't react to new things at all. it's trash only good for few things

Not even that user but that's not how AI works.
AI does that expensive training procedure a gorillion times, yes, but to generate one behavior tree that is top-fucking tier, you pop that shit into the game and it'll just follow that algorithm like any other. What AI does is DESIGN that tree, but to do that it needs feedback, and that needs a playable alpha/beta that can execute the algorithm with a player, that player can go from DSP to Donguri and whatever the player does the AI will take as feedback, rebalance the algorithm and try again. Then there's the issue that AI always aims to be better, so you might end up with an algorithm that makes the enemy win every time, balancing hijinks ensue and we spend more time training the fuckface than actually implementing it.
Basically you're right.

Why do games not use a difficulty that scales based on how good/shit you are?

Attached: 220px-God_Hand_(2006_Playstation_2)_video_game_cover_art.jpg (220x311, 28K)

Attached: 1096142219eeb5824d6829a4317ba3c5.png (1762x1710, 1.07M)

Maybe if the AIs in question were perfect AI. Perfect AI is extremely unlikely to ever happen though.
Maybe someone could train a neural network to demand rights.

Neural Networks operate on tensors of numbers, the game state is an array of structures not a big fucking list of numbers.

ML can be used for offline tasks like texture generation. You can also hand program an AI that will never lose without any sort of ML nonsense, the trick to creating a good AI is to make it hard but reasonable to overcome. There's no way to reason about what a neural network is going to do.

Don't tell me how AI works nigger, I work in this meme field.

So based on your image, you just want to have to do guardbreak a whole lot at the highest level difficulty?

FPBP

It isn't. Some do use neural networks Dota2 has Open AI which is beating pro players who makes millions a year. More companies don't because it's expensive.

>is beating pro players
hasn't happened in a real game of Dota 2. In a fake environment with lot's of stuff disabled, most of the heroes disabled, yeah.

It beats them under very limited circumstances with unfair advantages. The machine learns very simple tactics but it can execute these tactics flawlessly without error or any delay in execution. It also has more information about the game state than the player has at any given time.

You've been memed son. Give me a call when an AI can beat pokemon red

>twitch beat pokemon red
>neural network cant

Ya sure

what the fuck is this bullshit
fuck off nanashi, you used to be based but now you're just a fraud

You convert the relevant game state into reasonably normalized numbers and feed it in, I don't see what's so hard. You also don't feed the entire fucking game state into it, because that's way too much information and the AI isn't supposed to be omniscient anyway.

Look at the Dota 2 bots, they describe what the game state and action space looks like for the AI. Admittedly constructing that stuff and a decent reward function is a fuckton of work. It would only be worth it if a decent heuristic bot doesn't work well. But DESU a lot of strategy games give bullshit economy bonuses to their AI to create "higher difficulties" so the default state of bots is often bad

>decent reward function is a fuckton of work.
and that's understating it. The OpenAI bots didn't understand smokes/wards at all, probably because they couldn't figure out a good reward system for them. Sometimes well placed wards might not even do anything for you and you wouldn't know that until you analyze the game properly afterwards

I fucking want Nagatoro to fucking step on me FUCK

Machine learning takes a lot more work than telling it what to do because machines are fucking dumb.

AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

Attached: 1529954717771.jpg (1120x1732, 143K)

Why is there not enough porn of this girl? Do japs hate brown girls?

It cant. Not unless you use an algo that learns from the player input. But in that case it will just be mimicing the players run. Im talking about starting from scratch with reinforcement learning(what the dota bot is using). It cant be done currently because AI is fucking dumb.

I mean, the author is an actual porn artist known for femdom, and a lot of extremely fucked shit in general. It's shocking he got a very wholesome romance manga made out of his fetish waifu. Which granted, she was never a porn character, but she was definitely fetishistic.

Maybe artists just don't wanna step into the domain of an artist who created the character as an escape from drawing straight up porn.

Attached: 1529954654258.jpg (1109x1539, 193K)

>Why is there not enough porn of this girl?
because naughty tease is way hotter than any hardcore porn could ever be.

Attached: 1529248521366.png (960x1378, 940K)

>gee i wonder why a for-profit business would opt to the more efficient, less labor intensive option when trying to create a profitable product
wow gosh golly gee i don't know dave

>Do japs hate brown girls?
Are we already at the point where the old saying, "Little brown girls are the best!", is already forgotten ?

Attached: 1435189061719.png (1100x1200, 1.69M)

We're reaching smugness levels that shouldn't be possible

Attached: 1519807267397.jpg (292x187, 30K)

>we're never getting another Yupiel-sama

Attached: 1515201844876.gif (512x288, 859K)

God fucking dammit nanashi, you had it so right at first. Why did you have to turn it into harem garbage? I just wanted my lightly-disguised softcore porn.

buddy that was forgotten years ago

That's what happens when you don't take what's yours.

>nagatoro is still a thing
I remember years ago people would start randomly dumping that image set where she says she has a boyfriend already on Yea Forums which always caused people to chimp out and get ass-blasted. Fun threads.

Truly the worst timeline

Attached: 1506289791144.jpg (370x511, 96K)

That was the one time she went too fucking far. Ass-kicking is one thing but you don't destroy a man on the inside like that.
Maybe that's just me though, getting cheated on is a fucking horrible feeling.

Attached: 1529958196913.gif (802x1178, 2.83M)

people will repeatedly insist that FEAR's enemy AI is some of the best AI ever coded, even though it's really not any different from the marine squad AI from Half Life

the levels were just designed around the AI's capabilities, and the way the level design moved the players through the level and introduced enemies to the player meant that you felt like they were flanking you and strategizing. when in reality it was really just a bunch of map nodes dictating cover with a really basic state tree that had the soldiers shout out what you were doing

AI is not the thing that makes good enemies, good design makes good enemies. the STALKER AI is borderline retarded but the very basic interaction they have with the game world and the dynamic nature of it means they come off as a lot more human that you would think based on how they were scripted

i haven't heard that phrase since 'shoop da whoop'

It might be possible but even if it does you would need money to hire people to make your shit.

>even as a meme, she blueballs me
longest 5 minutes I've wasted

i-it can't be THAT old??
It would place it alongside with the DESU spam!

Attached: 1529247810242.png (1233x1858, 1.32M)

yes its that old. back when Yea Forums had a sidebar and moot was trolling people by telling them he was going to bring back WT Snacks

it's just not fucking fair

Attached: 1530067205545.png (381x538, 252K)

>tfw Chinese hour is so tempting but I have to get up for work in 5 hours

Attached: 1554334116763.png (497x480, 218K)

good thread

Attached: 1558382473197.gif (260x255, 38K)

He said AI Engineers, not wannabe losers who got a degree without actually knowing what they're doing.

>all I want is a bully gf

>every girl I date always wants to be choked and tossed around

Attached: 1531619401032.jpg (1280x720, 37K)

Please don't post these edits about my wife.

Thats what happens when you introduce a girl better than the heroine.