He could hallucinate death? Research by Google and Tel Aviv University has successfully simulated the game DOOM within a neural learning model called GameNGen.
It’s been a big year for the “Can It Run DOOM” scene. We’ve started DOOM Stool bacteriaAnd the knowledge of a mad scientist A Lab-grown rat brain To play death. But Google and Tel Aviv University’s research turned the scenario around GameNGen project— These researchers aren’t just running DOOM in an AI environment; simulation DOOM without using any traditional code, visual assets, or game engines. Figuratively speaking, we now have a neural learning model that can “think” about the existence of DOOM.
DOOM Simulated is instantly recognizable. It runs in full color at 20 frames per second and is powered by a single tensor processing unit (TPU). While this isn’t the first AI simulation of DOOM (nor can humans currently run it), it’s the most impressive and accurate yet, and it doesn’t require the latest hardware.
GameNGen training is achieved through a two-stage process. First, a reinforcement learning model (a reward-seeking AI, like a lab rat) was taught how to play DOOM. Her gaming sessions were recorded and passed to a diffusion model (an AI similar to the predictive text algorithm on your smartphone keyboard), which learned how to predict and create in-game visuals. The models were not exposed to DOOM’s source code or visual asset library.
“A complex video game, the popular one DOOM, can be run on a neural network (an enhanced version of the open Stable Diffusion version 1.4, in real time, while achieving visual quality similar to that of the original game). While not an exact simulation, the neural model is capable of Perform complex game state updates, such as calculating health and ammo, attacking enemies, damaging objects, opening doors, and continuing the game state over long paths.
While the AI DOOM simulation is clearly impressive, it’s not perfect. Many “complex game state updates,” such as a health meter or enemy movement, are influenced by the obvious visual elements we have come to associate with generative video. Objects turn into blurry blobs before snapping back into shape, sudden movement is often accompanied by smeared awkwardness, and the health meter fluctuates between numbers relentlessly. The AI is also unable to simulate any areas or functions in DOOM that were not explored during training.
However, GameNGen runs DOOM at better quality and frame rate than most PCs did in the mid-90s. And this is without the elegant DOOM engine (or any traditional game engine). Google research also found that when watching short After clips between 1.6 seconds and 3.2 seconds long, humans had great difficulty distinguishing between fake DOOM and real DOOM (their success rate was 58% to 60%). The simulated image is often ideal; He constantly fails to be perfect.
As for how this research will be used in the future, that’s anyone’s guess. Research by Google and Tel Aviv University has demonstrated that the interactive game can be run within the neural model paradigm. But they didn’t make a game from scratch. The arduous process of simulating a game within a neural model has no practical or economic benefit as of 2024, so GameNGen is little more than a proof of concept. It’s definitely not a product.
However, this research may lead to the development of a neural model that can be generated unique games. If generative game development can be achieved at a lower cost than traditional game development (while also providing an enjoyable experience for players), then something like GameNGen could become a viable product. But training may be the biggest hurdle here, as the AI will need to have a good understanding of how games work (GameNGen seems to rely heavily on visual feedback) and, more importantly, it will need a huge dataset with a variety of existing data. , Copyright games.
While I have done my best to explain this research, I suggest reading Diffusion models are real-time game engines White paper and visit GameNGen page on GitHub.
source: GameNGen
(tags for translation) Latest technology