Maia Chess

From Chessprogramming wiki
Jump to: navigation, search

Home * Engines * Maia Chess

Hermes and Maia [1]

Maia Chess, (Maia, Maiachess)
a chess engine featuring deep learning - as elaborated by their authors Reid McIlroy-Young, Siddhartha Sen, Jon Kleinberg, and Ashton Anderson in their research paper presented at the 26th ACM SIGKDD virtual conference in 2020 - with the aim to align superhuman AI with human behaviour [2], Russell Wang further joining the team. Like AlphaZero and Leela Chess Zero, Maia Chess uses a deep convolutional neural network (CNN) to predict moves. In contrast to the Zero training approaches of their inspirer, using reinforcement learning - Maia models are entirely trained by supervised learning, feeding in games of human players separated by 9 rating levels between 1100 and 1900 Elo. Further, Maia Chess only predicts moves by probing the net without any search. Maia chess is open source released under the terms under the GPL version 3, and consists of Python code relying on the scikit-learn library, along with various bash scripts. To play chess, Maia requires its models used in Lc0 [3] similar to any other Leela weights file - in UCI mode, nodes_1 needs to disable any search [4].

Model

Boards were represented as a 8×8×17 dimensional array with the 12 channels encoding pieces, 4 channels encoding castling rights, and one encoding whether the active player is white. The residual CNN has 6 residual blocks with two set of 2D CNNs with 64 channels and a 3×3 kernel [5].

Dedicated Maia

Maia Chess is incorporated inside modules of the Certabo Chessboard as chess playing engine, Certabo DaVinci [6] and Certabo Nano [7].


See also

Publications

Forum Posts

External Links

Chess Engine

Misc

feat.: Ticão Freitas, Luiz Otavio, Felipe Martins, Marcelo Martins, Enio Taquari

References

Up one level