25,161
edits
Changes
NNUE
,no edit summary
The neural network consists of four layers, W1 through W4. The input layer W1 is heavily overparametrized, feeding in the [[Board Representation|board representation]] for various king configurations.
The efficiency of the net is due to [[Incremental Updates|incremental update]] of W1 in [[Make Move|make]] and [[Unmake Move|unmake move]],
where only a fraction of its neurons need to be recalculated. The remaining three layers with 256x2x32-32x2x256, 32x32-and 32x1 neurons weights are computational less expensive, best calculated using appropriate [[SIMD and SWAR Techniques|SIMD instructions]] like [[AVX2]] on [[x86-64]], or if available, [[AVX-512]].
[[FILE:NNUE.jpg|none|border|text-bottom]]
=External Links=
* [https://github.com/yaneurao/YaneuraOu GitHub - yaneurao/YaneuraOu: YaneuraOu is the World's Strongest Shogi engine(AI player), WCSC29 1st winner, educational and USI compliant engine]
* [https://en.wikipedia.org/wiki/Efficiently_updatable_neural_network Efficiently updatable neural network | Wikipedia]
* [https://github.com/Tama4649/Kristallweizen/ GitHub - Tama4649/Kristallweizen: 第29回世界コンピュータ将棋選手権 準優勝のKristallweizenです。]
* [http://qhapaq.hatenablog.com/entry/2018/06/02/221612 次世代の将棋思考エンジン、NNUE関数を学ぼう（その１．ネットワーク構造編） - コンピュータ将棋 Qhapaq], June 02, 2018 (Japanese)
