MuZero

A model-based reinforcement learning algorithm that masters complex environments by learning its own internal dynamics instead of relying on pre-programmed rules.

Developed by Google DeepMind as the successor to AlphaZero, MuZero represents a breakthrough in general-purpose AI by eliminating the need for known environment dynamics. It uses a deep neural network to learn a latent model of its surroundings, focusing specifically on predicting three task-relevant quantities: reward, action policy, and state value. This approach allowed MuZero to match or exceed superhuman benchmarks across 57 different Atari games and classic board games like Go, Chess, and Shogi without ever being told the rules. Beyond gaming, the technology has transitioned into industrial applications: YouTube utilizes MuZero-based rate controllers to optimize video compression, achieving a 4% reduction in bitrate for VP9 video streams while maintaining visual quality.

https://deepmind.google/discover/blog/muzero-mastering-games-without-rules/