The model, called GameNGen, was developed by Dani Valevskis of Google Research and colleagues. According to the research paper, AI-generated Doom can be played for up to 20 seconds while retaining all the features of the original, such as game score points, ammo and map layout. Players can attack enemies, open doors, and interact with the environment quite normally.
Unfortunately, after the mentioned period, the model starts to run out of memory and the game crashes.
The original Doom game was released in 1993. and has since become a popular subject for computer science projects, including attempts to run it in unusual and limited hardware – such as toasters, treadmills and coffee machines.
However, in all these cases, the hardware is simply running the original game code. GameNGen works in a fundamentally different way: the artificial intelligence has learned by observation to recreate the game without seeing any of its programming code.
First, the researchers created an artificial intelligence model that learned to interact with Doom like a human. This model was then tasked with playing the game over and over, while a second AI model based on the Stable Diffusion image generator learned how hundreds of millions of inputs lead to changes in the state of the game.
The second model then essentially became a copy of the game, with all the knowledge, rules, and instructions from the original code encoded into an artificial neural network, its own architecture.
GameNGen developers say in their article that this is a proof of concept that games can be based not only on lines of programming code, but also on a neural network. They argue that games could be created from textual descriptions or concept drawings, making them cheaper to produce than using human programmers.
Andrew Rogoyski of the University of Surrey, UK, says the idea of making a neural network hallucinate a game’s environment and human interaction with it is an interesting step forward – but it won’t replace the people who make games.
“I don’t think this is the end of studying these games. I think game studios have the imagination, the skill to really create these worlds, to understand gameplay, to understand engagement, to understand how to draw us into the story. It’s not just nuts and bolts, bits and bytes, he says. “There’s something very human about creating an immersive experience that we humans enjoy and that currently or for the foreseeable future will mostly come from other people.”
The study is published on the pre-publication server arXiv.
Parengta pagal „New Scientist“.
2024-09-02 18:10:00