From 9b92ada935ddf920491156be22f609afaca4d840 Mon Sep 17 00:00:00 2001 From: Robert Nurnberg Date: Sun, 17 Mar 2024 15:39:01 +0100 Subject: [PATCH] Base WDL model on material count and normalize evals dynamically This PR proposes to change the parameter dependence of Stockfish's internal WDL model from full move counter to material count. In addition it ensures that an evaluation of 100 centipawns always corresponds to a 50% win probability at fishtest LTC, whereas for master this holds only at move number 32. See also https://github.com/official-stockfish/Stockfish/pull/4920 and the discussion therein. The new model was fitted based on about 340M positions extracted from 5.6M fishtest LTC games from the last three weeks, involving SF versions from e67cc979fd2c0e66dfc2b2f2daa0117458cfc462 (SF 16.1) to current master. The involved commands are for [WDL_model](https://github.com/official-stockfish/WDL_model) are: ``` ./updateWDL.sh --firstrev e67cc979fd2c0e66dfc2b2f2daa0117458cfc462 python scoreWDL.py updateWDL.json --plot save --pgnName update_material.png --momType "material" --momTarget 58 --materialMin 10 --modelFitting optimizeProbability ``` The anchor `58` for the material count value was chosen to be as close as possible to the observed average material count of fishtest LTC games at move 32 (`43`), while not changing the value of `NormalizeToPawnValue` compared to the move-based WDL model by more than 1. The patch only affects the displayed cp and wdl values. closes https://github.com/official-stockfish/Stockfish/pull/5121 No functional change --- src/evaluate.cpp | 4 +- src/nnue/nnue_misc.cpp | 18 ++++---- src/search.cpp | 7 +-- src/uci.cpp | 99 +++++++++++++++++++++++++----------------- src/uci.h | 6 +-- 5 files changed, 76 insertions(+), 58 deletions(-) diff --git a/src/evaluate.cpp b/src/evaluate.cpp index f4d18d8e..c7adf509 100644 --- a/src/evaluate.cpp +++ b/src/evaluate.cpp @@ -92,11 +92,11 @@ std::string Eval::trace(Position& pos, const Eval::NNUE::Networks& networks) { Value v = networks.big.evaluate(pos, false); v = pos.side_to_move() == WHITE ? v : -v; - ss << "NNUE evaluation " << 0.01 * UCI::to_cp(v) << " (white side)\n"; + ss << "NNUE evaluation " << 0.01 * UCI::to_cp(v, pos) << " (white side)\n"; v = evaluate(networks, pos, VALUE_ZERO); v = pos.side_to_move() == WHITE ? v : -v; - ss << "Final evaluation " << 0.01 * UCI::to_cp(v) << " (white side)"; + ss << "Final evaluation " << 0.01 * UCI::to_cp(v, pos) << " (white side)"; ss << " [with scaled NNUE, ...]"; ss << "\n"; diff --git a/src/nnue/nnue_misc.cpp b/src/nnue/nnue_misc.cpp index 7005a610..725d90d2 100644 --- a/src/nnue/nnue_misc.cpp +++ b/src/nnue/nnue_misc.cpp @@ -54,11 +54,11 @@ void hint_common_parent_position(const Position& pos, const Networks& networks) namespace { // Converts a Value into (centi)pawns and writes it in a buffer. // The buffer must have capacity for at least 5 chars. -void format_cp_compact(Value v, char* buffer) { +void format_cp_compact(Value v, char* buffer, const Position& pos) { buffer[0] = (v < 0 ? '-' : v > 0 ? '+' : ' '); - int cp = std::abs(UCI::to_cp(v)); + int cp = std::abs(UCI::to_cp(v, pos)); if (cp >= 10000) { buffer[1] = '0' + cp / 10000; @@ -90,9 +90,9 @@ void format_cp_compact(Value v, char* buffer) { // Converts a Value into pawns, always keeping two decimals -void format_cp_aligned_dot(Value v, std::stringstream& stream) { +void format_cp_aligned_dot(Value v, std::stringstream& stream, const Position& pos) { - const double pawns = std::abs(0.01 * UCI::to_cp(v)); + const double pawns = std::abs(0.01 * UCI::to_cp(v, pos)); stream << (v < 0 ? '-' : v > 0 ? '+' @@ -114,7 +114,7 @@ std::string trace(Position& pos, const Eval::NNUE::Networks& networks) { board[row][8 * 8 + 1] = '\0'; // A lambda to output one box of the board - auto writeSquare = [&board](File file, Rank rank, Piece pc, Value value) { + auto writeSquare = [&board, &pos](File file, Rank rank, Piece pc, Value value) { const int x = int(file) * 8; const int y = (7 - int(rank)) * 3; for (int i = 1; i < 8; ++i) @@ -125,7 +125,7 @@ std::string trace(Position& pos, const Eval::NNUE::Networks& networks) { if (pc != NO_PIECE) board[y + 1][x + 4] = PieceToChar[pc]; if (value != VALUE_NONE) - format_cp_compact(value, &board[y + 2][x + 2]); + format_cp_compact(value, &board[y + 2][x + 2], pos); }; // We estimate the value of each piece by doing a differential evaluation from @@ -180,13 +180,13 @@ std::string trace(Position& pos, const Eval::NNUE::Networks& networks) { { ss << "| " << bucket << " "; ss << " | "; - format_cp_aligned_dot(t.psqt[bucket], ss); + format_cp_aligned_dot(t.psqt[bucket], ss, pos); ss << " " << " | "; - format_cp_aligned_dot(t.positional[bucket], ss); + format_cp_aligned_dot(t.positional[bucket], ss, pos); ss << " " << " | "; - format_cp_aligned_dot(t.psqt[bucket] + t.positional[bucket], ss); + format_cp_aligned_dot(t.psqt[bucket] + t.positional[bucket], ss, pos); ss << " " << " |"; if (bucket == t.correctBucket) diff --git a/src/search.cpp b/src/search.cpp index fc92d1a9..9929ec27 100644 --- a/src/search.cpp +++ b/src/search.cpp @@ -155,7 +155,8 @@ void Search::Worker::start_searching() { { rootMoves.emplace_back(Move::none()); sync_cout << "info depth 0 score " - << UCI::value(rootPos.checkers() ? -VALUE_MATE : VALUE_DRAW) << sync_endl; + << UCI::to_score(rootPos.checkers() ? -VALUE_MATE : VALUE_DRAW, rootPos) + << sync_endl; } else { @@ -1898,10 +1899,10 @@ std::string SearchManager::pv(const Search::Worker& worker, ss << "info" << " depth " << d << " seldepth " << rootMoves[i].selDepth << " multipv " << i + 1 - << " score " << UCI::value(v); + << " score " << UCI::to_score(v, pos); if (worker.options["UCI_ShowWDL"]) - ss << UCI::wdl(v, pos.game_ply()); + ss << UCI::wdl(v, pos); if (i == pvIdx && !tb && updated) // tablebase- and previous-scores are exact ss << (rootMoves[i].scoreLowerbound diff --git a/src/uci.cpp b/src/uci.cpp index cf0e3f09..cc03005f 100644 --- a/src/uci.cpp +++ b/src/uci.cpp @@ -28,6 +28,7 @@ #include #include #include +#include #include #include "benchmark.h" @@ -44,9 +45,8 @@ namespace Stockfish { -constexpr auto StartFEN = "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1"; -constexpr int NormalizeToPawnValue = 356; -constexpr int MaxHashMB = Is64Bit ? 33554432 : 2048; +constexpr auto StartFEN = "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1"; +constexpr int MaxHashMB = Is64Bit ? 33554432 : 2048; namespace NN = Eval::NNUE; @@ -338,15 +338,43 @@ void UCI::position(Position& pos, std::istringstream& is, StateListPtr& states) } } -int UCI::to_cp(Value v) { return 100 * v / NormalizeToPawnValue; } +namespace { +std::pair win_rate_params(const Position& pos) { -std::string UCI::value(Value v) { + int material = pos.count() + 3 * pos.count() + 3 * pos.count() + + 5 * pos.count() + 9 * pos.count(); + + // The fitted model only uses data for material counts in [10, 78], and is anchored at count 58. + double m = std::clamp(material, 10, 78) / 58.0; + + // Return a = p_a(material) and b = p_b(material), see github.com/official-stockfish/WDL_model + constexpr double as[] = {-185.71965483, 504.85014385, -438.58295743, 474.04604627}; + constexpr double bs[] = {89.23542728, -137.02141296, 73.28669021, 47.53376190}; + + 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]; + + return {a, b}; +} + +// The win rate model is 1 / (1 + exp((a - eval) / b)), where a = p_a(material) and b = p_b(material). +// It fits the LTC fishtest statistics rather accurately. +int win_rate_model(Value v, const Position& pos) { + + auto [a, b] = win_rate_params(pos); + + // Return the win rate in per mille units, rounded to the nearest integer. + return int(0.5 + 1000 / (1 + std::exp((a - double(v)) / b))); +} +} + +std::string UCI::to_score(Value v, const Position& pos) { assert(-VALUE_INFINITE < v && v < VALUE_INFINITE); std::stringstream ss; if (std::abs(v) < VALUE_TB_WIN_IN_MAX_PLY) - ss << "cp " << to_cp(v); + ss << "cp " << to_cp(v, pos); else if (std::abs(v) <= VALUE_TB) { const int ply = VALUE_TB - std::abs(v); // recompute ss->ply @@ -358,6 +386,30 @@ std::string UCI::value(Value v) { return ss.str(); } +// Turns a Value to an integer centipawn number, +// without treatment of mate and similar special scores. +int UCI::to_cp(Value v, const Position& pos) { + + // In general, the score can be defined via the the WDL as + // (log(1/L - 1) - log(1/W - 1)) / ((log(1/L - 1) + log(1/W - 1)) + // Based on our win_rate_model, this simply yields v / a. + + auto [a, b] = win_rate_params(pos); + + return std::round(100 * int(v) / a); +} + +std::string UCI::wdl(Value v, const Position& pos) { + std::stringstream ss; + + int wdl_w = win_rate_model(v, pos); + int wdl_l = win_rate_model(-v, pos); + int wdl_d = 1000 - wdl_w - wdl_l; + ss << " wdl " << wdl_w << " " << wdl_d << " " << wdl_l; + + return ss.str(); +} + std::string UCI::square(Square s) { return std::string{char('a' + file_of(s)), char('1' + rank_of(s))}; } @@ -383,41 +435,6 @@ std::string UCI::move(Move m, bool chess960) { return move; } -namespace { -// The win rate model returns the probability of winning (in per mille units) given an -// eval and a game ply. It fits the LTC fishtest statistics rather accurately. -int win_rate_model(Value v, int ply) { - - // The fitted model only uses data for moves in [8, 120], and is anchored at move 32. - double m = std::clamp(ply / 2 + 1, 8, 120) / 32.0; - - // 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. - constexpr double as[] = {-1.06249702, 7.42016937, 0.89425629, 348.60356174}; - constexpr double bs[] = {-5.33122190, 39.57831533, -90.84473771, 123.40620748}; - - // Enforce that NormalizeToPawnValue corresponds to a 50% win rate at move 32. - static_assert(NormalizeToPawnValue == int(0.5 + 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]; - - // Return the win rate in per mille units, rounded to the nearest integer. - return int(0.5 + 1000 / (1 + std::exp((a - double(v)) / b))); -} -} - -std::string UCI::wdl(Value v, int ply) { - std::stringstream ss; - - int wdl_w = win_rate_model(v, ply); - int wdl_l = win_rate_model(-v, ply); - int wdl_d = 1000 - wdl_w - wdl_l; - ss << " wdl " << wdl_w << " " << wdl_d << " " << wdl_l; - - return ss.str(); -} Move UCI::to_move(const Position& pos, std::string& str) { if (str.length() == 5) diff --git a/src/uci.h b/src/uci.h index dd55862a..237928d9 100644 --- a/src/uci.h +++ b/src/uci.h @@ -42,11 +42,11 @@ class UCI { void loop(); - static int to_cp(Value v); - static std::string value(Value v); + static int to_cp(Value v, const Position& pos); + static std::string to_score(Value v, const Position& pos); static std::string square(Square s); static std::string move(Move m, bool chess960); - static std::string wdl(Value v, int ply); + static std::string wdl(Value v, const Position& pos); static Move to_move(const Position& pos, std::string& str); static Search::LimitsType parse_limits(const Position& pos, std::istream& is);