Efficiently updatable neural network

This article is part of the series on
Chess programming
Board representations 0x88 Bitboards
Evaluation functions Deep neural networks (Transformers) Attention Efficiently updatable neural networks Handcrafted evaluation functions Piece-square tables Reinforcement learning Stochastic gradient descent Supervised learning Texel tuning Unsupervised learning
Graph and tree search algorithms Minimax Alpha-beta pruning Principal variation search Quiescence search Monte Carlo tree search
Chess computers Belle ChessMachine ChipTest Cray Blitz Deep Blue Deep Thought HiTech Hydra Mephisto Saitek
Chess engines AlphaZero Chess Tiger Crafty CuckooChess Deep Fritz Dragon by Komodo Chess Fairy-Max Fritz Fruit GNU Chess HIARCS Houdini Ikarus Junior KnightCap Komodo Leela Chess Zero MChess Pro Mittens MuZero Naum REBEL Rybka Shredder Sjeng SmarThink Stockfish Torch Turochamp Zappa
v t e

An efficiently updatable neural network (NNUE, a Japanese wordplay on Nue, sometimes stylised as ƎUИИ) is a neural network-based evaluation function whose inputs are piece-square tables, or variants thereof like the king-piece-square table.^[1] NNUE is used primarily for the leaf nodes of the alpha–beta tree.^[2] While being slower than handcrafted evaluation functions, NNUE does not suffer from the 'blindness beyond the current move' problem.^[3]

NNUE was invented by Yu Nasu and introduced to computer shogi in 2018.^[4][5] On 6 August 2020, NNUE was for the first time ported to a chess engine, Stockfish 12.^[6][7] Since 2021, many of the top rated classical chess engines such as Komodo Dragon have an NNUE implementation to remain competitive.

NNUE runs efficiently on central processing units (CPU) without a requirement for a graphics processing unit (GPU). In contrast, deep neural network-based chess engines such as Leela Chess Zero require a GPU.

The neural network used for the original 2018 computer shogi implementation consists of four weight layers: W1 (16-bit integers) and W2, W3 and W4 (8-bit). It has 4 fully-connected layers, ReLU activation functions, and outputs a single number, being the score of the board.

W1 encoded the king's position and therefore this layer needed only to be re-evaluated once the king moved. It used incremental computation and single instruction multiple data (SIMD) techniques along with appropriate intrinsic instructions.^[4]