MuZero

https://en.wikipedia.org/wiki/MuZero

Training compute (board games: 16 TPUs training, 1000 for self-play, 800 simulations per step; Atari: 8 and 32 TPUs, 50 simulations); matched AlphaZero in Go after about 1M steps; poor on 6 Atari games; successors EfficientZero and Stochastic MuZero.

Self-play planning with learned models (AlphaZero, MuZero, Dreamer)