Difference between revisions of "NNUE"

From Chessprogramming wiki
Jump to: navigation, search
Line 87: Line 87:
 
* [http://www.talkchess.com/forum3/viewtopic.php?f=2&t=75433 Seer] by [[Gerd Isenberg]], [[CCC]], October 18, 2020 » [[Seer]]
 
* [http://www.talkchess.com/forum3/viewtopic.php?f=2&t=75433 Seer] by [[Gerd Isenberg]], [[CCC]], October 18, 2020 » [[Seer]]
 
* [http://www.talkchess.com/forum3/viewtopic.php?f=2&t=75482 BBC 1.3 + Stockfish NNUE released!] by [[Maksim Korzh]], [[CCC]], October 21, 2020 » [[BBC]]
 
* [http://www.talkchess.com/forum3/viewtopic.php?f=2&t=75482 BBC 1.3 + Stockfish NNUE released!] by [[Maksim Korzh]], [[CCC]], October 21, 2020 » [[BBC]]
* [http://www.talkchess.com/forum3/viewtopic.php?f=2&t=75500 Mayhem NNUE - New NN engine] by [[Toni Helminen|JohnWoe]], [[CCC]], October 22, 2020 » [[Mayhem NNUE]]
+
* [http://www.talkchess.com/forum3/viewtopic.php?f=2&t=75500 Mayhem NNUE - New NN engine] by [[Toni Helminen|JohnWoe]], [[CCC]], October 22, 2020 » [[Mayhem]]
 
* [http://www.talkchess.com/forum3/viewtopic.php?f=7&t=75501 Centipawns vs Millipawns with NNUE] by Madeleine Birchfield, [[CCC]], October 23, 2020 » [[Centipawns]], [[Millipawns]]
 
* [http://www.talkchess.com/forum3/viewtopic.php?f=7&t=75501 Centipawns vs Millipawns with NNUE] by Madeleine Birchfield, [[CCC]], October 23, 2020 » [[Centipawns]], [[Millipawns]]
 
* [http://www.talkchess.com/forum3/viewtopic.php?f=7&t=75506 NNUE Question - King Placements] by [[Andrew Grant]], [[CCC]], October 23, 2020 » [[Stockfish NNUE#NNUE Structure|Stockfish NNUE - NNUE Structure]]
 
* [http://www.talkchess.com/forum3/viewtopic.php?f=7&t=75506 NNUE Question - King Placements] by [[Andrew Grant]], [[CCC]], October 23, 2020 » [[Stockfish NNUE#NNUE Structure|Stockfish NNUE - NNUE Structure]]

Revision as of 08:42, 18 June 2021

Home * Learning * Neural Networks * NNUE

Toriyama Sekien - Nue (鵺) [1] [2]

NNUE, (ƎUИИ Efficiently Updatable Neural Networks)
a Neural Network architecture intended to replace the evaluation of Shogi, chess and other board game playing alpha-beta searchers running on a CPU. NNUE was introduced in 2018 by Yu Nasu [3], and was used in Shogi adaptations of Stockfish such as YaneuraOu [4], and Kristallweizen [5], apparently with AlphaZero strength [6].

Stockfish NNUE

As reported by Henk Drost [7], Nodchip incorporated NNUE into the chess playing Stockfish 10 as a proof of concept. Stockfish NNUE was born, and in summer 2020 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, becoming stronger than its original [8], finally responsible for the huge strength improvement of Stockfish 12.

NNUE Engines

see Category:NNUE

Being tempted by the success of Stockfish NNUE and attracted by how easy the method and small the code is, many engine developers have started testing and applying NNUE. For quick trials and evaluating before going into serious development, some of them borrowed and/or rewrote NNUE code and use networks from Stockfish NNUE. Most of them reported positive results, such as David Carteau with Orion [9], Ehsan Rashid with DON [10], various Stockfish derivatives by Michael Byrne [11], and Volodymyr Shcherbyna with Igel [12] using the Night Nurse NNUE net by Dietrich Kappe [13]. Daniel Shawul added NNUE support à la CFish into his egbbdll probing library of Scorpio [14] [15], making it even easier to use NNUE. The promising engines Halogen 7 and 8 by Kieren Pearson, and Seer by Connor McMonigle came with their own, distinct NNUE implementations, and on November 10, 2020, the commercial Dragon by Komodo Chess aka Komodo NNUE appeared [16], trying to close the gap to Stockfish NNUE. The commercial Fat Fritz 2.0, based on a slightly modified Stockfish 12 using a customized, double sized network, was released by ChessBase in February 2021.

NN Structure

The neural network of Stockfish NNUE consists of four layers, W1 through W4. The input layer W1 is heavily overparametrized, feeding in the board representation for various king configurations. The efficiency of the net is due to incremental update of W1 in make and unmake move, where only a fraction of its neurons need to be recalculated. The remaining three layers with 32x2x256, 32x32 and 32x1 weights are computational less expensive, best calculated using appropriate SIMD instructions like AVX2 on x86-64, or if available, AVX-512.

NNUE.jpg

NNUE structure with incremental update [17]

See also

Publications

Forum Posts

2020

January ...

Re: The Stockfish of shogi by Gian-Carlo Pascutto, CCC, January 18, 2020

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

September

Re: First success with neural nets by Jonathan Kreuzer, CCC, November 11, 2020 » Checkers

October

Re: Hacking around CFish NNUE by Daniel Shawul, CCC, October 15, 2020 » Scorpio
Re: How to scale stockfish NNUE score? by Daniel Shawul, CCC, October 17, 2020

November

Re: Speculations about NNUE development (was New engine releases 2020) by Connor McMonigle, CCC, November 12, 2020
Re: Speculations about NNUE development (was New engine releases 2020) by Connor McMonigle, CCC, November 12, 2020 » Dragon by Komodo Chess, Seer, Halogen

December

2021

January

February

March

April

May

June

Re: Commercial Release of Ethereal 13.00 (NNUE) for AVX2 Systems by Andrew Grant, CCC, June 04, 2021 » HalfKP
Re: will Tcec allow Stockfish with a Leela net to play? by Daniel Shawul, CCC, June 18, 2021 » Scorpio
Re: will Tcec allow Stockfish with a Leela net to play? by Vivien Clauzon, CCC, June 18, 2021 » Minic

External Links

NNUE

Learn Next Generation Shogi Thinking Engine, NNUE Function (Part 1. Network Structure) - Computer Shogi
Let's Learn Next Generation Shogi Thinking Engine, NNUE Function (Part 2. Remodeling/Learning) - Computer Shogi

NNUE libraries

Some developers disintegrate and rewrite the Stockfish NNUE code into independent libraries which can be much easier to embed into other chess engines.

Source Code

Misc

References

  1. Nue (鵺) from the Konjaku Gazu Zoku Hyakki (今昔画図続百鬼) by Toriyama Sekien, circa 1779, Wikimedia Commons
  2. Re: What does NNUE actually mean by Henk Drost, CCC, July 29, 2020
  3. Yu Nasu (2018). ƎUИИ Efficiently Updatable Neural-Network based Evaluation Functions for Computer Shogi. Ziosoft Computer Shogi Club, pdf (Japanese with English abstract)
  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. The Stockfish of shogi by Larry Kaufman, CCC, January 07, 2020
  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. Orion 0.7 : NNUE experiment by David Carteau, CCC, August 19, 2020
  10. Re: New engine releases 2020...Don NNUE 2020? by supersharp77, CCC, August 19, 2020
  11. ... the last shall be first ... by MikeB, CCC, 19 Aug 2020
  12. Introducing Igel chess engine by Volodymyr Shcherbyna, CCC, 20 Aug 2020
  13. Night Nurse 0.2 by Dietrich Kappe, CCC, August 19, 2020
  14. Re: Hacking around CFish NNUE by Daniel Shawul, CCC, October 15, 2020
  15. Re: How to scale stockfish NNUE score? by Daniel Shawul, CCC, October 17, 2020
  16. Dragon by Komodo Chess by Larry Kaufman, CCC, November 10, 2020
  17. Image from Yu Nasu (2018). ƎUИИ Efficiently Updatable Neural-Network based Evaluation Functions for Computer Shogi. Ziosoft Computer Shogi Club, pdf (Japanese with English abstract)
  18. Translation of Yu Nasu's NNUE paper by Dominik Klein, CCC, January 07, 2021
  19. Outer product from Wikipedia
  20. Tensor product from Wikipedia
  21. PyTorch from Wikipedia
  22. TensorFlow from Wikipedia
  23. Vafra by Robert Jurjević
  24. nnue-trainer by Jon Dart, CCC, March 27, 2021

Up one Level