deepcrayon
/
stockfish
1
0
Fork 0
Code Releases Activity

Fallback to NNUE 

If the classical eval ends up much smaller than estimated fall back to NNUE.
Also use multiply instead of divide for the threshold comparison for smoother transitions without rounding.

STC https://tests.stockfishchess.org/tests/view/5f3a5011b38d442594aabdfe
LLR: 2.96 (-2.94,2.94) {-0.50,1.50}
Total: 57352 W: 6325 L: 6135 D: 44892
Ptnml(0-2): 277, 4748, 18482, 4846, 323

LTC https://tests.stockfishchess.org/tests/view/5f3aee9db38d442594aabe82
LLR: 2.95 (-2.94,2.94) {0.25,1.75}
Total: 16232 W: 897 L: 781 D: 14554
Ptnml(0-2): 19, 679, 6616, 771, 31

closes https://github.com/official-stockfish/Stockfish/pull/3023

bench: 4026216

-----

Recommended net: https://tests.stockfishchess.org/api/nn/nn-82215d0fd0df.nnue
pull/3024/head
mstembera 2020-08-16 15:23:50 -07:00 committed by Joost VandeVondele
parent 581b92e4a7
commit 1bcc981a5a
1 changed files with 6 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
src/evaluate.cpp
View File

@ -114,7 +114,8 @@ namespace {
constexpr Value LazyThreshold1 = Value(1400);
constexpr Value LazyThreshold2 = Value(1300);
constexpr Value SpaceThreshold = Value(12222);
constexpr Value NNUEThreshold = Value(575);
constexpr Value NNUEThreshold1 = Value(550);
constexpr Value NNUEThreshold2 = Value(150);
// KingAttackWeights[PieceType] contains king attack weights by piece type
constexpr int KingAttackWeights[PIECE_TYPE_NB] = { 0, 0, 81, 52, 44, 10 };
@ -939,10 +940,13 @@ make_v:
Value Eval::evaluate(const Position& pos) {
bool classical = !Eval::useNNUE
|| abs(eg_value(pos.psq_score())) >= NNUEThreshold * (16 + pos.rule50_count()) / 16;
|| abs(eg_value(pos.psq_score())) * 16 > NNUEThreshold1 * (16 + pos.rule50_count());
Value v = classical ? Evaluation<NO_TRACE>(pos).value()
: NNUE::evaluate(pos) * 5 / 4 + Tempo;
if (classical && Eval::useNNUE && abs(v) * 16 < NNUEThreshold2 * (16 + pos.rule50_count()))
v = NNUE::evaluate(pos) * 5 / 4 + Tempo;
// Damp down the evaluation linearly when shuffling
v = v * (100 - pos.rule50_count()) / 100;