It seems to me like dynamic enemy placement in a premade level, smart procgen and aesthetically pleasing random enemy design shuffle is an obvious choice, but what else could possibly be created using this method?
Picrelated, a hirame slut colorized using a neural network algorithm.
Basically neural network algorithms mimic the way human learning works by reinforcing connections between "neurons" (of course they're not actually neurons they're just there to mimic the way it works in the human brain)
So such an algorithm can be trained through trial and error to mimic and generate Baroque music from scratch, or be taught how to perfectly speedrun Mario
Cooper Rodriguez
hnnngg
Christopher Garcia
I fail to see where it mentions using NNA
Joseph Diaz
I want some user to go and colorize a doujin by Hirame fully
Jeremiah Hughes
That sounds awesome, custom tailor made perfect videogames for everyone!
Ethan Stewart
Sounds like a great way to bloat your budget and ensure you'll spend years debugging, patching and balancing. Maybe, maybe people like David Braben could find a way to use it to its full potential but there are more efficient ways to get similar results.
Camden Garcia
I would imagine NNs would be used in an offline capacity, not dynamically Although maybe you could use a NN to find the best parameters to the functions you are going to call in the actual program.
As for what to use them on…how would you use a NN for enemy and level design without existing data to generate patterns from? Using data from other games would lead to suboptimal results no matter how much data you feed into it, since the mechanics of every game is subtly different.
In order to steer that kind of high level design using algorithms (NN or otherwise), I think we'd have to wait until we get other AIs that can competently play a game (like, without requiring a major research effort, something devs can just plug into their game and let it learn for a weekend.). Then you could have player AIs as your playtesters, who would generate various play statistics for you as you changed the level design parameters. That's a vague an far off vision, but it's more realistic than trying to use NNs in the near term for level design.
There's probably already NNs being used in a few games right now for tiny artistic things though.
Camden Wood
The game itself is built around them. I guess it doesn't say it anywhere on the store page, but when trying to mod the game, it explains it in more detail
But it is driven by a NNAI
Unfortunately, it's also a very linear, easy to win "game" with few actual choices, there's almost always an optimal option for the player to make
Ryan Price
This is a highly fetishized feature among roguelike developers. The "virtual GM". There are a few very highly skilled people with years of practice in failing to design your example.
Carter Morris
I would do a platformer. First they train the AI before shipping just enough to actually know how to move around and jump. The "procreational goal" or generational goal of the AI is to catch the player. If you need one statistic, it would be the time it takes for the AI to catch the player, and the AI has all the same platform/walljump/parkour maneuvres as the player. Every time the player plays, the AI learns. There would also be a game function (At the main menu) that allows it to use a randomly chosen selection of your played games (Game records your movement and creates a ghost for AI to chase) to run either 100 or 1000 generations of the AI.
There is no victory in the game, you just run around whatever level playing keep away from the AI, which would continually better at navigating and catching you based on how you personally play.
I guess kind of like N+ but instead of having timed life it's however long it takes for you to screw up and get caught.
If you wanted more mechanics you could add environmental dangers, enemies, or turrets. A win is counted for you if you get the AI killed, and a win for the AI if you die. The AI would come pre-trained to not walk into these.
A bit niche, but you asked.
Adam Rodriguez
I want to kill Hirame and eat his heart to inherit his drawing style
too bad nigger magic isn't a real thing
Tyler Moore
Targeted ingame advertisement.
Kevin James
fuck
Luke Cox
delete this
John Ramirez
I recognize this from somewhere but I need the sauce.
Hunter Rodriguez
Search for hirame on exhentai
Noah Sullivan
The crops are a lot more erotic than the actual thing if you ask me.
Grayson Wilson
[Service Heaven (Karei,Turtle.Fish.Paint)] UnLove S
Jeremiah Turner
I would ban you
Christian Clark
a little help never hurt anyone.
Nolan Allen
...
Noah Davis
user, the artist's name was dropped in OP's post. Spoonfeeding is one thing, this is regurgitating pre-digested sauce.
Aiden Baker
Woah, she fucks a dog and then she fucks a robot?! What a dirty whore!
Xavier Rogers
I can think of at least one near-term use, but modern devs refuse to have decent bots for whatever reason.
Chase Reyes
A deep mind-like AI that can learn through ML how to play a competitive game at the highest level will do wonders for game designers trying to balance it. There is one that plays Melee and its already giving top players a hard time.
The help of those would make fightans much better at release.
Andrew Rivera
That's pretty clever idea. Make a novel game designed around having a learning opponent instead of asking where can a NN be shoehorned into the production of a regular game.
Parker Murphy
I really want proper believable NPC AI based on behaviors created through NNA
Levi Lee
SWAT 4 already did that in a pretty effective way.
Googles Deepmind was supposed to get SC2 account but it is Blizzard and dead game.
It wouldn't be so bad if they would actually fucking use it. God fucking damn you Blizzard and Google.
Zachary Perry
You know why patents barely exist in video games? Software patents are easy as fuck to get around, even down to changing how you phrase things. Nintendo already did the DNN shit with Super Smash Bros for the Wii U, since you got a bot that was suppose to learn off the people it was playing against.
That patent will probably expire before someone actually gets off their ass to innovate that isn't an indie dev, unless it's used like how marketing firms use it to screw with people. The video game industry is literally as stagnant as Hollywood right now.
DNN sounds like it'd be shit in games anyways
Jaxon Cook
Neural networks are a fad. How would you even efficiently train for half of the stuff you propose? As funny as is, it's probably the most realistic use.
Eli Gray
I was even less bothered than you think. Google was one of few companies that had money to toy with any form of AI and they were only one interested in actually using it in vidya. Them having working patent meant nothing because from current perspective they were only one doing it anyway. However I must agree if patents would work in the sightless I wouldn't be able to trace all Dota characters to their W3 counterparts.
Anyways Blizzard just announced that SC1 HD is happening. There is still chance for Deepmind to change its name to Avilo and find a way to shit on people so hard they quit 90% of matches. Google said it will learn from people after all.
Landon Peterson
maybe
Daniel Hernandez
Neural Networks are very, very slow unless you use several cores or PCs with them. Their application for singleplayer games would be very minimal unless there was only a very small amount of variables and things to learn.
For instance, there's that game where you "train" robots by giving them points according to how close they can reach certain goals like shooting the enemy and moving towards the flag. It takes several generations before they can even walk straight but the end result is mostly the same for every player. You'll only use them when they can rush wherever you place your banner and shoot anyone nearby while keeping some distance.
There's also the problem that, since their only use is to teach AI, then a better AI is the only product you can make with them, so you'd need a game that can make use of it. So the natural conclusion: AI for enemies and similar elements for multiplayer games.
For instance, imagine if AI for arena shooters was connected online and they'd share their learnings every match. The larger the amount of players, the faster it would learn and the more varied it's learning would be to the point that you'd get AI that could compete with the best players since it learned not just with them but with everyone else. This makes full use of it and solves the speed problem since every player is essentially a core.
That's a basic use, a better more interesting one would be the AI Director in Left4Dead, who actually does very little things. If the item spawn was dynamic up to it's location, the Director could actually pick different places like under tables or behind doors to spawn items if it wanted to increase difficulty, learning the places where players don't look, as well as spawning Special Zombies based on location and how hard it is fighting against one in a particular place versus a different special, all based on past fights.
Alexander King
I'm fairly certain For Honor uses AI training algorithms to justify being always online even in singleplayer. Maybe look into that, not sure what it exactly implements.
Honestly though, I don't see a use for AI training outside of niche areas, much in the same I don't see procedural generation being all that useful. AI training currently requires a large sample size to actually learn from, and the changes through it most likely definitely won't be felt in single play session and most likely not among sessions in close time proximity. The only way I see it being effective is in multiplayer or co-op experiences that are meant to be revisited repeatedly and thus require some dynamic elements to keep it fresh, possibly in MMOs too. I don't see much use in single-player games that would benefit more from being 100% hand-crafted and intentionally designed to provide a better experience overall.
Tyler Hernandez
The biggest problem with making AI with neural networks is how to train it, if you want an AI that plays against the player you first need a standard AI in place and many, many, many session in which the AI can observe and learn, most oftenly they will only become relevant when the player already got bored with the game.
So to even start training they use 50 emulators and a super computer.
Took a whole day of super computing with massive paralelization just so it can fight the game AI and not get wrecked.
Two weeks of training to get it to be playable against some pros.
And this is only on ONE AI on ONE map
And it still manages to kill itself in confusion when it faces some unpredicted situtation.
So, can you guys finally understand how impossible it is to properly train an AI to make it "learn" how to play against the player? Unless it's some kind of online fighting game that the AI can collect and analyse matches and learn from thousands of players playing countless times, you are not gonna get AI based on neural networks, our dreams of "organic" and "evolving" NPCs are misguided, ANNs are not gonna do this, there could be other ways, but not ANNs. Though I think it's ironic how dota has the best chance of realizing our dream of ANN AI than any other game.
Henry Edwards
imagine if you could put deep learning AI on NPCs so you would get Tay.ai tier companions instead of 3 premade responses every time
Ayden Garcia
You mean like Super Smash Bros Melee
All the problems they describe can be fixed and the results are promising. I don't know how you got "impossible" out of it rather than "eventual".