Nvidia Teaches Neural Network to Recreate Pacman

[InsertPi]

Original link: https://www.theverge.com/2020/5/22/21266251/nvidia-ai-gamegan-recreate-pac-man-virutal-environment

 

TLDR:

 

Quote

Nvidia researchers taught an AI system to recreate the game of Pac-Man simply by watching it being played.

There’s no coding involved, no pre-rendered images for the software to draw on. The AI model is simply fed visual data of the game in action along with the accompanying controller inputs and then recreates it frame by frame from this information. The resulting game is playable by humans, and Nvidia says it will be releasing it online in the near future.

 

Quote

Nvidia says work like this shows how artificial intelligence will be used for game design in the future. Developers can input their work into the AI and use it to create variations or maybe design new levels. “You could use this to mash different games together,” Sanja Fidler, director of Nvidia’s Toronto research lab, told journalists, “giving additional power to games developers by [letting them] blend together different games.”

 

Quote

Fidler told journalists that to recreate Pac-Man, GameGAN had to be trained on some 50,000 episodes. Getting that gameplay data from humans wasn’t feasible, so the team used an AI agent to generate the data. Unfortunately, the AI agent was so good at the game that it hardly ever died.

“That made it hard for the AI trying to recreate the game to learn the concept of dying,” says Fidler. Instead, in early versions of the AI-generated Pac-Man, GameGAN tweaked the game so that ghosts never actually reached the title character but trail directly behind it like baby ducks following a parent.

 

Some gameplay is previewed here:

 

 

This is a pretty nice leap forward with generative adversarial networks (the type of neural network used in this project), and is yet another testament to how they can generate rich, complex results. As Sanja Fidler said in the article, this could be beneficial for game developers. Potentially, it could mean that GANs could be used to create games out of a handful of mechanics. Just feed in a bunch of game mechanics to a GAN and let it generate a game that incorporates all those mechanics!

 

Perhaps in the future this could be used for making games out of TV shows or movies with long enough screen times, by feeding it a show or movie, along with what the button inputs would possibly look like for a finished game. Or rather, using a neural network as a realistic physics engine for a game by feeding it video of objects colliding or otherwise interacting. The technology isn't quite there yet, but this is a big leap towards a future where things like that are possible.

[Froody129]

What I'm really excited for is AI AI. Most racing games suck if they're not online, and many FPS games have awful AI that just make it a rinse-and-repeat exercise.

 

This is an awesome demo. Imagine if you got a bunch of soldiers and had them run airsoft games where they play like they would in a combat situation. If you get enough data from them you could use that to teach an AI to think and react like them. 

[StDragon]

Maybe we can get it to learn Wargames (DEFCON)

 

"Greetings professor Falken. A strange game. The only winning move is not to play."

[Arika]

I mean AI is cool and all, but no amount of machine learning will ever beat the marvel that is Twitch plays Pokemon which just goes to show that anything is possible through random chance and absolute chaos

[RejZoR]

I have doubts. They keep on saying how Ai just makes shit out of thin air. People sometimes don't understand what the fuck they are doing with games and some Ai will just "understand" what's going on and code the game by observing it without actually coding it. Sorry, but it just sounds like the biggest nonsense in the world. Game or program code doesn't just miraculously appear out of nowhere while not being there, while being there because we obviously have a game here. Game or a program needs to have a very specific sequence of code in order to even execute and not just crash endlessly or simply not do anything. Secondly, this "Ai" had to "know" it's gonna be doing a game. Because if you would feed this wonder Ai Excel spreadsheet instead of a game I very much doubt it would just create a working spreadsheet editor without generating any code. So, someone predefined what it has to be generating. And we're back at the "generating" shit. It has to be a highly specific and exact code to work. It just all sounds and reads like total absolute nonsense garbage. Literally only way this could ever work even slightly is that you have a game designing framework in place, you tell the "Ai" you're monitoring a game to design a game and it "learns" from what it sees and recreates the game in the game making framework (like Games Factory, but where machine learning has ability to generate assets as required, that's literally only way you're not coding a thing because framework does that, you just tell it what, how and where). It can't just make shit up out of nothing and have a functioning code without having one. Like, dafaq is wrong with this world, like everyone forgot how coding shit works and it's not like you just wave a magic wand and everything just pops into existence somehow.

 

Also, the title should be "NVIDIA used machine learning to visually copy a game through a pre-existing game framework". But we all know such title wouldn't fly with all the clickbait rules, so it's "NVIDIA's Ai created a game". Which it didn't. At least not out of thin air into a working game without knowing what it's doing and having environment (framework) to do it. I guess this will eventually simplify game design because when you'll have a FRAMEWORK for game building in place that will be able to process and prepare game actions and happenings into a functioning code underneath without developer having to code it by hand, it'll take less time to make complex games. Then again, game FRAMEWORKS already do this. Just replace "framework" with "game engine". Unreal Engine is known to have things within its framework super simplified and developers often don't have to touch or write any code, they navigate through it via friendly visual controls that then output clean expected code for game to work. Sure you still need a human, but it's whole lot easier to click and stack few parameters to make a shader instead of coding it from scratch and hoping you don't make a typo. And you'll be having har time making Excel spreadsheet editor with it because the framework (lets say Unreal Engine) just isn't primarily designed for this. The same applies to this magical "Ai" from NVIDIA. It's not a sentient Ai that can make shit out of nothing, it had things prepared and set in place to visually copy a game via "learning". "Ai" is such a dumb buzzword...

[StDragon]

6 minutes ago, RejZoR said:

"Ai" is such a dumb buzzword...

No doubt about it, that AI has been pillaged of its meaning for profit.

 

I suppose even toilets have AI in them. They have a single sensor that "learns" when I walk up to it and leave before it flushes on my behalf. How intelligent! How artificial! 

[akaREMO]

10 hours ago, Arika S said:

I mean AI is cool and all, but no amount of machine learning will ever beat the marvel that is Twitch plays Pokemon which just goes to show that anything is possible through random chance and absolute chaos

Twitch Plays is absolutely hilarious.