MuZero: plan in a learned model

AlphaZero-style search calls the true game rules at every node of its tree. MuZero deletes that call and learns the simulator instead: h encodes the observation, g predicts the next latent state, f judges positions. The same search then runs inside the learned model. This module rebuilds MuZero small enough to read in a sitting, trains it on your machine, and measures where the learned model is right and where it drifts.

What you'll learn: learned dynamics · latent-space MCTS · the value equivalence principle · K-step unroll training · the self-play flywheel & Reanalyze · honest failure modes.

Code companion: nanoworld. Prerequisite: the AlphaZero course. Part of the world-models course on tuul.dev.