This is Muse, Microsoft’s new AI capable of generating virtual worlds for gameplay and preserving classic video games

Microsoft has presented Muse, a new artificial intelligence system (AI) capable of generating graphics and controller actions for video games. This development is qualified as the first human and world action model (WHAM) and aims to become an assistance tool for the creatives of the world Gamer.

The product is the result of a collaboration between Microsoft Research and Ninja Theory, one of the studies First Party of Xbox. In general terms, Muse can predict images, simulate entries and produce game sequences based on the physics and operation of pre -existing games. Engineers point out, in a Published article In the magazine Naturethat this system integrates a series of space-time models that allow it to understand the dynamics of current video games. This capacity arises from a data -based training technique obtained from more than 500,000 human game sessions.

The Microsoft team compiled more than seven years of continuous and anonymous games of Bleeding Edgea 3D multiplayer title developed by Ninja Theory. The collection covers the seven game maps and accumulates more than 100,000 million images of games, matched with the corresponding control actions. As a result, the “7 Maps DATASET“, An information base used to train the most extensive version of the What.

This methodology has allowed Muse to learn directly from the games, instead of following predefined rules such as other AI models. In such a way, It can generate realistic game images and precise control actions without pre -established knowledge. The system operates with 1.6 billion parameters and a context length of a second. Each frame is encoded in 540 Tokens with a resolution of 300 × 180 pixels.

Muse’s performance was evaluated based on three aspects: consistency, diversity and persistence. “The consistency refers to the model of the model to generate sequences that respect the dynamics of the game. Diversity measures its ability to create variants from the same initial indication, reflecting different ways in which the game could evolve. Persistence evaluates its ability to incorporate and maintain user modifications in the generated sequences, as a character that is copied and attached to an image of the game, ”explains Microsoft.