Add NNUE evaluation
authornodchip <nodchip@gmail.com>
Wed, 5 Aug 2020 15:11:15 +0000 (17:11 +0200)
committerJoost VandeVondele <Joost.VandeVondele@gmail.com>
Thu, 6 Aug 2020 14:37:45 +0000 (16:37 +0200)
commit84f3e867903f62480c33243dd0ecbffd342796fc
tree3827a1afc1a587d4e72d0d9877de6c9a7baf72f8
parent9587eeeb5ed29f834d4f956b92e0e732877c47a7
Add NNUE evaluation

This patch ports the efficiently updatable neural network (NNUE) evaluation to Stockfish.

Both the NNUE and the classical evaluations are available, and can be used to
assign a value to a position that is later used in alpha-beta (PVS) search to find the
best move. The classical evaluation computes this value as a function of various chess
concepts, handcrafted by experts, tested and tuned using fishtest. The NNUE evaluation
computes this value with a neural network based on basic inputs. The network is optimized
and trained on the evalutions of millions of positions at moderate search depth.

The NNUE evaluation was first introduced in shogi, and ported to Stockfish afterward.
It can be evaluated efficiently on CPUs, and exploits the fact that only parts
of the neural network need to be updated after a typical chess move.
[The nodchip repository](https://github.com/nodchip/Stockfish) provides additional
tools to train and develop the NNUE networks.

This patch is the result of contributions of various authors, from various communities,
including: nodchip, ynasu87, yaneurao (initial port and NNUE authors), domschl, FireFather,
rqs, xXH4CKST3RXx, tttak, zz4032, joergoster, mstembera, nguyenpham, erbsenzaehler,
dorzechowski, and vondele.

This new evaluation needed various changes to fishtest and the corresponding infrastructure,
for which tomtor, ppigazzini, noobpwnftw, daylen, and vondele are gratefully acknowledged.

The first networks have been provided by gekkehenker and sergiovieri, with the latter
net (nn-97f742aaefcd.nnue) being the current default.

The evaluation function can be selected at run time with the `Use NNUE` (true/false) UCI option,
provided the `EvalFile` option points the the network file (depending on the GUI, with full path).

The performance of the NNUE evaluation relative to the classical evaluation depends somewhat on
the hardware, and is expected to improve quickly, but is currently on > 80 Elo on fishtest:

60000 @ 10+0.1 th 1
https://tests.stockfishchess.org/tests/view/5f28fe6ea5abc164f05e4c4c
ELO: 92.77 +-2.1 (95%) LOS: 100.0%
Total: 60000 W: 24193 L: 8543 D: 27264
Ptnml(0-2): 609, 3850, 9708, 10948, 4885

40000 @ 20+0.2 th 8
https://tests.stockfishchess.org/tests/view/5f290229a5abc164f05e4c58
ELO: 89.47 +-2.0 (95%) LOS: 100.0%
Total: 40000 W: 12756 L: 2677 D: 24567
Ptnml(0-2): 74, 1583, 8550, 7776, 2017

At the same time, the impact on the classical evaluation remains minimal, causing no significant
regression:

sprt @ 10+0.1 th 1
https://tests.stockfishchess.org/tests/view/5f2906a2a5abc164f05e4c5b
LLR: 2.94 (-2.94,2.94) {-6.00,-4.00}
Total: 34936 W: 6502 L: 6825 D: 21609
Ptnml(0-2): 571, 4082, 8434, 3861, 520

sprt @ 60+0.6 th 1
https://tests.stockfishchess.org/tests/view/5f2906cfa5abc164f05e4c5d
LLR: 2.93 (-2.94,2.94) {-6.00,-4.00}
Total: 10088 W: 1232 L: 1265 D: 7591
Ptnml(0-2): 49, 914, 3170, 843, 68

The needed networks can be found at https://tests.stockfishchess.org/nns
It is recommended to use the default one as indicated by the `EvalFile` UCI option.

Guidelines for testing new nets can be found at
https://github.com/glinscott/fishtest/wiki/Creating-my-first-test#nnue-net-tests

Integration has been discussed in various issues:
https://github.com/official-stockfish/Stockfish/issues/2823
https://github.com/official-stockfish/Stockfish/issues/2728

The integration branch will be closed after the merge:
https://github.com/official-stockfish/Stockfish/pull/2825
https://github.com/official-stockfish/Stockfish/tree/nnue-player-wip

closes https://github.com/official-stockfish/Stockfish/pull/2912

This will be an exciting time for computer chess, looking forward to seeing the evolution of
this approach.

Bench: 4746616
59 files changed:
.travis.yml
AUTHORS
README.md [moved from Readme.md with 79% similarity]
appveyor.yml
src/Makefile
src/benchmark.cpp
src/bitbase.cpp
src/bitboard.cpp
src/bitboard.h
src/endgame.cpp
src/endgame.h
src/evaluate.cpp
src/evaluate.h
src/main.cpp
src/material.cpp
src/material.h
src/misc.cpp
src/misc.h
src/movegen.cpp
src/movegen.h
src/movepick.cpp
src/movepick.h
src/nnue/architectures/halfkp_256x2-32-32.h [new file with mode: 0644]
src/nnue/evaluate_nnue.cpp [new file with mode: 0644]
src/nnue/evaluate_nnue.h [new file with mode: 0644]
src/nnue/features/feature_set.h [new file with mode: 0644]
src/nnue/features/features_common.h [new file with mode: 0644]
src/nnue/features/half_kp.cpp [new file with mode: 0644]
src/nnue/features/half_kp.h [new file with mode: 0644]
src/nnue/features/index_list.h [new file with mode: 0644]
src/nnue/layers/affine_transform.h [new file with mode: 0644]
src/nnue/layers/clipped_relu.h [new file with mode: 0644]
src/nnue/layers/input_slice.h [new file with mode: 0644]
src/nnue/nnue_accumulator.h [new file with mode: 0644]
src/nnue/nnue_architecture.h [new file with mode: 0644]
src/nnue/nnue_common.h [new file with mode: 0644]
src/nnue/nnue_feature_transformer.h [new file with mode: 0644]
src/pawns.cpp
src/pawns.h
src/position.cpp
src/position.h
src/psqt.cpp
src/search.cpp
src/search.h
src/syzygy/tbprobe.cpp
src/syzygy/tbprobe.h
src/thread.cpp
src/thread.h
src/thread_win32_osx.h
src/timeman.cpp
src/timeman.h
src/tt.cpp
src/tt.h
src/tune.cpp
src/tune.h
src/types.h
src/uci.cpp
src/uci.h
src/ucioption.cpp