Algorithm Median Mean Frames IMP ALA

Atari games, best results are highlighted in bold . For per game scores, see Table 10. Hyper-parameters were tuned per game. All other parameters are held constant. All experiments in this paper are based on a JAX (Bradbury et al., 2018) implementation of MuZero, For experiments in environments with continuous actions, we used the extension to MuZero proposed in (Hubert et al., 2021).