Stockfish NNUE

From Chessprogramming wiki
Revision as of 23:58, 19 August 2020 by Phhnguyen (talk | contribs) (Elo Progress with NNUE)
Jump to: navigation, search

Home * Engines * Stockfish * NNUE

Stockfish NNUE Logo [1]

Stockfish NNUE,
a Stockfish branch by Hisayori Noda aka Nodchip, which uses Efficiently Updatable Neural Networks - stylized as ƎUИИ or reversed as NNUE - to replace its standard evaluation. NNUE, introduced in 2018 by Yu Nasu [2], were previously successfully applied in Shogi evaluation functions embedded in a Stockfish based search [3], such as YaneuraOu [4], and Kristallweizen [5]. In 2019, Nodchip incorporated NNUE into Stockfish 10 - as a proof of concept, and with the intention to give something back to the Stockfish community [6]. After support and announcements by Henk Drost in May 2020 [7] and subsequent enhancements, Stockfish NNUE was established and recognized. In summer 2020, with more people involved in testing and training, the computer chess community bursts out enthusiastically due to its rapidly raising playing strength with different networks trained using a mixture of supervised and reinforcement learning methods. Despite the approximately halved search speed, Stockfish NNUE seemingly became stronger than its original [8]. In August 2020, Fishtest revealed Stockfish NNUE was stronger than the classical one at least 80 Elo[9]. In July 2020, the playing code of NNUE was put into the official Stockfish repository as a branch for further development and examination. In August that playing code merged to the master branch and become an official part of the engine. However, the training code still remained in Nodchip's repository [10] [11].

NNUE Structure

The neural network consists of four layers. The input layer is heavily overparametrized, feeding in the board representation for all king placements per side. The efficiency of NNUE is due to incremental update of the input layer outputs in make and unmake move, where only a tiny fraction of its neurons need to be considered [12]. The remaining three layers with 2x256x32, 32x32 and 32x1 weights are computational less expensive, hidden layer 1 and 2 with ReLu activation [13] [14], best calculated using appropriate SIMD instructions performing fast 8-bit/16-bit integer vector arithmetic, like MMX, SSE2 or AVX2 on x86/x86-64, or, if available, AVX-512.

StockfishNNUELayers.png

NNUE layers in action [15]

Nodchip explained on Discord on June 2020 [16]:

41,024 = 64 * 641. 64 comes from the number of the cells where king may exist. 641 = 64 * 5 * 2 + 1. 64 here comes from the number of the cells where a piece other than king may exist. 5 is the number of piece types other than king. 2 is the number of the colors, white and black. 1 is a captured piece.
"+ 1" is BONA_PIECE_ZERO.
Here "bona" means "bonanza" which is a popular computer shogi engine. It introduced the feature "p" for the first time.
BonaPieces are contained in the evalList. It is updated by Position::do_move() and Position::undo_move(), and used by NNUE to calculate the network parameters between the input layer and the first hidden layer.
About the calculation, the following text will be helpful. This text is sent to RocketMiningPoo on Twitter.
"We add the i-th COLUMN of the W{0} to the z{0} for each i, where the i-th element is set to 1. And we subtract the i-th COULMN of the W{0} from the z{0} for each i, where the i-th element is set to 0. This operation is "accumulate" in your question." may be right... I hope that someone will double check.

Networks

All networks are built by some volunteers but not by any big community (differs from Leela Chess Zero) and can be uploaded into Fishtest for testing. Networks with good test results will be released officially on Fishtest website [17].

Hybrid

In August 2020 a new patch changed Stockfish NNUE into a hybrid engine: it uses NNUE evaluation only on quite balanced material positions, otherwise uses the classical one. It could speed up to 10% and gain 20 Elo [18].

Strong Points

For Users

  • Runs with CPU only, and doesn't require expensive video cards, as well the need for installing drivers and 3rd specific libraries. Thus it is much easier to install (compared to engines using deep convolutional neural networks, such as Leela Chess Zero) and suitable for almost all modern computers. Using a GPU is even not practical for that small net - host-device-latency aka. kernel-launch-overhead [19] [20] to a then underemployed GPU are not sufficient for the intended NPS range [21]
  • Releases with only one network (via UCI options), that help to delete users' confusion from finding, selecting and setting up. The network is selected carefully from Fishtest

For Developers

  • Reuses and gets benefits from the very optimized search function of Stockfish as well as almost all Stockfish's code
  • Requires much smaller training sets. Some high score networks can be built with the effort of one or a few people within a few days. It doesn't require the massive computing from a supercomputer and/or from community
  • Relative simple for computer chess programmers as well as being encouraged by some impressive successes, many developers joined or continued to work. The Official Stockfish repository shows the numbers of commits, ideas increased significantly after merging NNUE.

Elo Progress with NNUE

Fishtest team did a Regression Tests in August 2020, exclusively from 40.000 games tests at long time control (60sec + 0.6 for single core, 30sec + 0.3 for 8 cores tests). The result was used to create the graph of Stockfish progress since version 11 [22].

Nnueprogress.png

See also

Forum Posts

2020 ...

January ...

July

Re: NNUE accessible explanation by Jonathan Rosenthal, CCC, July 23, 2020
Re: NNUE accessible explanation by Jonathan Rosenthal, CCC, July 24, 2020
Re: NNUE accessible explanation by Jonathan Rosenthal, CCC, August 03, 2020

August

Re: this will be the merge of a lifetime : SF 80 Elo+ by Henk Drost, CCC, August 04, 2020

External Links

Basics

Source

Networks

Binaries

Rating Lists

Misc

References

  1. Stockfish NNUE Logo from GitHub - nodchip/Stockfish: UCI chess engine by Nodchip
  2. Yu Nasu (2018). ƎUИИ Efficiently Updatable Neural-Network based Evaluation Functions for Computer Shogi. Ziosoft Computer Shogi Club, pdf (Japanese with English abstract)
  3. The Stockfish of shogi by Larry Kaufman, CCC, January 07, 2020
  4. GitHub - yaneurao/YaneuraOu: YaneuraOu is the World's Strongest Shogi engine(AI player), WCSC29 1st winner, educational and USI compliant engine
  5. 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。
  6. Stockfish NNUE – The Complete Guide, June 19, 2020 (Japanese and English)
  7. Stockfish NN release (NNUE) by Henk Drost, CCC, May 31, 2020
  8. Can the sardine! NNUE clobbers SF by Henk Drost, CCC, July 16, 2020
  9. Introducing NNUE Evaluation, August 06, 2020
  10. NNUE merge · Issue #2823 · official-stockfish/Stockfish · GitHub by Joost VandeVondele, July 25, 2020
  11. GitHub - nodchip/Stockfish: UCI chess engine by Nodchip
  12. Re: NNUE accessible explanation by Jonathan Rosenthal, CCC, July 23, 2020
  13. Stockfish/halfkp_256x2-32-32.h at master · official-stockfish/Stockfish · GitHub
  14. Stockfish/clipped_relu.h at master · official-stockfish/Stockfish · GitHub
  15. Image courtesy Roman Zhukov, revised version of the image posted in Re: Stockfish NN release (NNUE) by Roman Zhukov, CCC, June 17, 2020
  16. Don't understand NNUE by Lucasart, CCC, August 14, 2020
  17. Neural Net download and statistics
  18. NNUE evaluation threshold by MJZ1977 · Pull Request #2916 · official-stockfish/Stockfish · GitHub, August 06, 2020
  19. AB search with NN on GPU... by Srdja Matovic, CCC, August 13, 2020 » GPU
  20. kernel launch latency - CUDA / CUDA Programming and Performance - NVIDIA Developer Forums by LukeCuda, June 18, 2018
  21. stockfish with graphics card by h1a8, CCC, August 06, 2020
  22. https://github.com/glinscott/fishtest/wiki/Regression-Tests Regression Tests, August 2020
  23. An info by Sylwy, CCC, July 25, 2020
  24. You can now look inside NNUE and look at its Per square value estimation by Henk Drost, CCC, August 04, 2020

Up one Level