From 200cd29d7fefe3ced2ea83eb1b35b993ddecb635 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Ste=CC=81phane=20Nicolet?= Date: Fri, 19 Nov 2021 20:04:35 +0100 Subject: [PATCH] Introduce Optimism Current master implements a scaling of the raw NNUE output value with a formula equivalent to 'eval = alpha * NNUE_output', where the scale factor alpha varies between 1.8 (for early middle game) and 0.9 (for pure endgames). This feature allows Stockfish to keep material on the board when she thinks she has the advantage, and to seek exchanges and simplifications when she thinks she has to defend. This patch slightly offsets the turning point between these two strategies, by adding to Stockfish's evaluation a small "optimism" value before actually doing the scaling. The effect is that SF will play a little bit more risky, trying to keep the tension a little bit longer when she is defending, and keeping even more material on the board when she has an advantage. We note that this patch is similar in spirit to the old "Contempt" idea we used to have in classical Stockfish, but this implementation differs in two key points: a) it has been tested as an Elo-gainer against master; b) the values output by the search are not changed on average by the implementation (in other words, the optimism value changes the tension/exchange strategy, but a displayed value of 1.0 pawn has the same signification before and after the patch). See the old comment https://github.com/official-stockfish/Stockfish/pull/1361#issuecomment-359165141 for some images illustrating the ideas. ------- finished yellow at STC: LLR: -2.94 (-2.94,2.94) <0.00,2.50> Total: 165048 W: 41705 L: 41611 D: 81732 Ptnml(0-2): 565, 18959, 43245, 19327, 428 https://tests.stockfishchess.org/tests/view/61942a3dcd645dc8291c876b passed LTC: LLR: 2.95 (-2.94,2.94) <0.50,3.00> Total: 121656 W: 30762 L: 30287 D: 60607 Ptnml(0-2): 87, 12558, 35032, 13095, 56 https://tests.stockfishchess.org/tests/view/61962c58cd645dc8291c8877 ------- How to continue from there? a) the shape (slope and amplitude) of the sigmoid used to compute the optimism value could be tweaked to try to gain more Elo, so the parameters of the sigmoid function in line 391 of search.cpp could be tuned with SPSA. Manual tweaking is also possible using this Desmos page: https://www.desmos.com/calculator/jhh83sqq92 b) in a similar vein, with two recents patches affecting the scaling of the NNUE evaluation in evaluate.cpp, now could be a good time to try a round of SPSA tuning of the NNUE network; c) this patch will tend to keep tension in middlegame a little bit longer, so any patch improving the defensive aspect of play via search extensions in risky, tactical positions would be welcome. ------- closes https://github.com/official-stockfish/Stockfish/pull/3797 Bench: 6184852 --- src/evaluate.cpp | 12 ++++++++---- src/misc.h | 27 +++++++++++++++++++++++++++ src/search.cpp | 15 ++++++++++----- src/thread.h | 1 + 4 files changed, 46 insertions(+), 9 deletions(-) diff --git a/src/evaluate.cpp b/src/evaluate.cpp index 9fdadbb74f3..c140a7e43ce 100644 --- a/src/evaluate.cpp +++ b/src/evaluate.cpp @@ -1091,11 +1091,15 @@ Value Eval::evaluate(const Position& pos) { v = Evaluation(pos).value(); // classical else { - int scale = 898 - + 24 * pos.count() - + 33 * pos.non_pawn_material() / 1024; + int scale = 898 + + 24 * pos.count() + + 33 * pos.non_pawn_material() / 1024; - v = NNUE::evaluate(pos, true) * scale / 1024; // NNUE + Value nnue = NNUE::evaluate(pos, true); // NNUE + Color stm = pos.side_to_move(); + Value optimism = pos.this_thread()->optimism[stm]; + + v = (nnue + optimism) * scale / 1024 - optimism; if (pos.is_chess960()) v += fix_FRC(pos); diff --git a/src/misc.h b/src/misc.h index 718e5558178..bc2f8bf6e41 100644 --- a/src/misc.h +++ b/src/misc.h @@ -138,6 +138,33 @@ class ValueList { std::size_t size_ = 0; }; + +// sigmoid(t, x0, y0, C, P, Q) implements a sigmoid-like function +// using only integers, with the following properties: +// +// - sigmoid is centered in (x0, y0) +// - sigmoid has amplitude [-P/Q , P/Q] instead of [-1 , +1] +// - limit is (y0 - P/Q) when t tends to -infinity +// - limit is (y0 + P/Q) when t tends to +infinity +// - the slope can be adjusted using C > 0, smaller C giving a steeper sigmoid +// - the slope of the sigmoid when t = x0 is P/(Q*C) +// - sigmoid is increasing with t when P > 0 and Q > 0 +// - to get a decreasing sigmoid, call with -t, or change sign of P +// - mean value of the sigmoid is y0 +// +// Use to draw the sigmoid + +inline int64_t sigmoid(int64_t t, int64_t x0, + int64_t y0, + int64_t C, + int64_t P, + int64_t Q) +{ + assert(C > 0); + return y0 + P * (t-x0) / (Q * (abs(t-x0) + C)) ; +} + + /// xorshift64star Pseudo-Random Number Generator /// This class is based on original code written and dedicated /// to the public domain by Sebastiano Vigna (2014). diff --git a/src/search.cpp b/src/search.cpp index fabb0ff07ac..fa2f506bc1e 100644 --- a/src/search.cpp +++ b/src/search.cpp @@ -334,8 +334,10 @@ void Thread::search() { nodesLastExplosive = nodes; nodesLastNormal = nodes; - state = EXPLOSION_NONE; - trend = SCORE_ZERO; + state = EXPLOSION_NONE; + trend = SCORE_ZERO; + optimism[ us] = Value(25); + optimism[~us] = -optimism[us]; int searchAgainCounter = 0; @@ -381,11 +383,14 @@ void Thread::search() { alpha = std::max(prev - delta,-VALUE_INFINITE); beta = std::min(prev + delta, VALUE_INFINITE); - // Adjust trend based on root move's previousScore (dynamic contempt) - int tr = 113 * prev / (abs(prev) + 147); - + // Adjust trend and optimism based on root move's previousScore + int tr = sigmoid(prev, 0, 0, 147, 113, 1); trend = (us == WHITE ? make_score(tr, tr / 2) : -make_score(tr, tr / 2)); + + int opt = sigmoid(prev, 0, 25, 147, 113, 2); + optimism[ us] = Value(opt); + optimism[~us] = -optimism[us]; } // Start with a small aspiration window and, in the case of a fail diff --git a/src/thread.h b/src/thread.h index 387937390c1..cd206faab98 100644 --- a/src/thread.h +++ b/src/thread.h @@ -68,6 +68,7 @@ class Thread { int selDepth, nmpMinPly; Color nmpColor; ExplosionState state; + Value optimism[COLOR_NB]; Position rootPos; StateInfo rootState;