Changes

Jump to: navigation, search

Stockfish NNUE

1,067 bytes removed, 22:05, 28 August 2020
Isozaki seems not NNUE inventor, I don't like the redundancies and vague statements in "Ideas around", Srdja's post already mentioned.
=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 <ref>[https://github.com/official-stockfish/Stockfish/pull/2916 NNUE evaluation threshold by MJZ1977 · Pull Request #2916 · official-stockfish/Stockfish · GitHub], August 06, 2020</ref>. At that point, NNUE helped to increase already around 100 Elo for Stockfish. In the same month, Stockfish changed the default mode of using evaluation functions from classic to hybrid one, the last step to completely accept NNUE.
 
=Ideas around=
Based on NNUE techniques, the community has discussed some ideas or even tried some of them:
* Use NNUE code and Stockfish NNUE networks for other engines
* Use Video cards in hope to speed up NNUE evaluation <ref>[http://www.talkchess.com/forum3/viewtopic.php?f=2&t=74702 stockfish with graphics card] by h1a8, [[CCC]], August 06, 2020</ref>. However, with current small networks and [[Incremental Updates|incremental update]] people guessed the GPU may not help much. Another topic <ref>[http://talkchess.com/forum3/viewtopic.php?f=7&t=74771 AB search with NN on GPU...] by Srdja Matovic, [[CCC]], August 13, 2020</ref> studied that the latencies between CPU and GPU become too large for the way calling NNUE evaluation without batches, may limit much the maximum speed
* Use other engines to train networks for Stockfish NNUE<ref>[http://talkchess.com/forum3/viewtopic.php?f=2&t=74510 LizardFish 0.1 NNUE now available] by Dietrich Kappe, [[CCC]], August 20, 2020</ref>
=Strong Points=
Anyway, a fix from rotate to flip has to be done from producer and consumer sides, and is likely to void some training sessions.
=Suggestions from an NNUE inventor=In reply to [[Unai Corzo]], [[Motohiro Isozaki]] aka Yaneurao<ref>[https://github.com/yaneurao GitHub yaneurao]</ref>, suggested 3 techniques that applied successfully to Shogi and can be brought back to Stockfish NNUE and may improve it another 100 - 200 Elo <ref>[http://yaneuraou.yaneu.com/2020/08/21/3-technologies-in-shogi-ai-that-could-be-used-for-chess-ai/ 3 technologies in shogi AI that could be used for chess AI], [[Motohiro Isozaki]], August 2020</ref> <ref>[https://github.com/official-stockfish/Stockfish/issues/2915#issuecomment-678112885, GitHub - NNUE ideas and discussion (post-merge). #2915], August 2020</ref>:
* Optimizing all parameters together by [https://en.wikipedia.org/wiki/Stochastic_optimization stochastic optimization]
* Switching between multi-evaluation functions, according to [[Game Phases|game phases]]
Retrieved from "https://www.chessprogramming.org/Special:MobileDiff/21032"

Navigation menu