From ad2aa8c06f438de8b8bb7b7c8726430e3f2a5685 Mon Sep 17 00:00:00 2001 From: Joost VandeVondele Date: Mon, 31 Oct 2022 20:36:43 +0100 Subject: [PATCH] Normalize evaluation Normalizes the internal value as reported by evaluate or search to the UCI centipawn result used in output. This value is derived from the win_rate_model() such that Stockfish outputs an advantage of "100 centipawns" for a position if the engine has a 50% probability to win from this position in selfplay at fishtest LTC time control. The reason to introduce this normalization is that our evaluation is, since NNUE, no longer related to the classical parameter PawnValueEg (=208). This leads to the current evaluation changing quite a bit from release to release, for example, the eval needed to have 50% win probability at fishtest LTC (in cp and internal Value): June 2020 : 113cp (237) June 2021 : 115cp (240) April 2022 : 134cp (279) July 2022 : 167cp (348) With this patch, a 100cp advantage will have a fixed interpretation, i.e. a 50% win chance. To keep this value steady, it will be needed to update the win_rate_model() from time to time, based on fishtest data. This analysis can be performed with a set of scripts currently available at https://github.com/vondele/WLD_model fixes https://github.com/official-stockfish/Stockfish/issues/4155 closes https://github.com/official-stockfish/Stockfish/pull/4216 No functional change --- src/nnue/evaluate_nnue.cpp | 4 ++-- src/uci.cpp | 12 ++++++++---- src/uci.h | 7 +++++++ 3 files changed, 17 insertions(+), 6 deletions(-) diff --git a/src/nnue/evaluate_nnue.cpp b/src/nnue/evaluate_nnue.cpp index ba2ed367..4715fed0 100644 --- a/src/nnue/evaluate_nnue.cpp +++ b/src/nnue/evaluate_nnue.cpp @@ -220,7 +220,7 @@ namespace Stockfish::Eval::NNUE { buffer[0] = (v < 0 ? '-' : v > 0 ? '+' : ' '); - int cp = std::abs(100 * v / PawnValueEg); + int cp = std::abs(100 * v / UCI::NormalizeToPawnValue); if (cp >= 10000) { buffer[1] = '0' + cp / 10000; cp %= 10000; @@ -251,7 +251,7 @@ namespace Stockfish::Eval::NNUE { buffer[0] = (v < 0 ? '-' : v > 0 ? '+' : ' '); - double cp = 1.0 * std::abs(int(v)) / PawnValueEg; + double cp = 1.0 * std::abs(int(v)) / UCI::NormalizeToPawnValue; sprintf(&buffer[1], "%6.2f", cp); } diff --git a/src/uci.cpp b/src/uci.cpp index d5e2c2c3..19e2b0cb 100644 --- a/src/uci.cpp +++ b/src/uci.cpp @@ -207,13 +207,17 @@ namespace { // The coefficients of a third-order polynomial fit is based on the fishtest data // for two parameters that need to transform eval to the argument of a logistic // function. - double as[] = { 0.50379905, -4.12755858, 18.95487051, 152.00733652}; - double bs[] = {-1.71790378, 10.71543602, -17.05515898, 41.15680404}; + constexpr double as[] = { 1.04790516, -8.58534089, 39.42615625, 316.17524816}; + constexpr double bs[] = { -3.57324784, 22.28816201, -35.47480551, 85.60617701 }; + + // Enforce that NormalizeToPawnValue corresponds to a 50% win rate at ply 64 + static_assert(UCI::NormalizeToPawnValue == int(as[0] + as[1] + as[2] + as[3])); + double a = (((as[0] * m + as[1]) * m + as[2]) * m) + as[3]; double b = (((bs[0] * m + bs[1]) * m + bs[2]) * m) + bs[3]; // Transform the eval to centipawns with limited range - double x = std::clamp(double(100 * v) / PawnValueEg, -2000.0, 2000.0); + double x = std::clamp(double(v), -4000.0, 4000.0); // Return the win rate in per mille units rounded to the nearest value return int(0.5 + 1000 / (1 + std::exp((a - x) / b))); @@ -312,7 +316,7 @@ string UCI::value(Value v) { stringstream ss; if (abs(v) < VALUE_MATE_IN_MAX_PLY) - ss << "cp " << v * 100 / PawnValueEg; + ss << "cp " << v * 100 / NormalizeToPawnValue; else ss << "mate " << (v > 0 ? VALUE_MATE - v + 1 : -VALUE_MATE - v) / 2; diff --git a/src/uci.h b/src/uci.h index 76a893f9..f5f2c385 100644 --- a/src/uci.h +++ b/src/uci.h @@ -30,6 +30,13 @@ class Position; namespace UCI { +// Normalizes the internal value as reported by evaluate or search +// to the UCI centipawn result used in output. This value is derived from +// the win_rate_model() such that Stockfish outputs an advantage of +// "100 centipawns" for a position if the engine has a 50% probability to win +// from this position in selfplay at fishtest LTC time control. +const int NormalizeToPawnValue = 348; + class Option; /// Define a custom comparator, because the UCI options should be case-insensitive