MiniZero: Comparative Analysis of AlphaZero and MuZero on Go, Othello, and Atari Games

arXiv:2310.11305v3 Announce Type: replace-cross
Abstract: This paper presents MiniZero, a zero-knowledge learning framework that supports four state-of-the-art algorithms, including AlphaZero, MuZero, Gumbel AlphaZero, and Gumbel MuZero. While these algorithms have demonstrated super-human performance in many games, it remains unclear which among them is most suitable or efficient for specific tasks. Through MiniZero, we systematically evaluate the performance of each algorithm in two board games, 9x9 Go and 8x8 Othello, as well as 57 Atari games. For two board games, using …

alphazero analysis arxiv atari games comparative analysis cs.ai cs.lg games type