Use winning_percentage_wdl in learn

This commit is contained in:
tttak
2020-08-24 22:56:08 +09:00
parent 7ee8a2bbb7
commit 4ce30d9522
3 changed files with 80 additions and 21 deletions

View File

@@ -238,27 +238,34 @@ namespace {
// The win rate model returns the probability (per mille) of winning given an eval
// and a game-ply. The model fits rather accurately the LTC fishtest statistics.
int win_rate_model(Value v, int ply) {
// The model captures only up to 240 plies, so limit input (and rescale)
double m = std::min(240, ply) / 64.0;
// Coefficients of a 3rd order polynomial fit based on fishtest data
// for two parameters needed to transform eval to the argument of a
// logistic function.
double as[] = {-8.24404295, 64.23892342, -95.73056462, 153.86478679};
double bs[] = {-3.37154371, 28.44489198, -56.67657741, 72.05858751};
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 eval to centipawns with limited range
double x = Utility::clamp(double(100 * v) / PawnValueEg, -1000.0, 1000.0);
// Return win rate in per mille (rounded to nearest)
return int(0.5 + 1000 / (1 + std::exp((a - x) / b)));
return int(0.5 + UCI::win_rate_model_double(v, ply));
}
} // namespace
// The win rate model returns the probability (per mille) of winning given an eval
// and a game-ply. The model fits rather accurately the LTC fishtest statistics.
double UCI::win_rate_model_double(double v, int ply) {
// The model captures only up to 240 plies, so limit input (and rescale)
double m = std::min(240, ply) / 64.0;
// Coefficients of a 3rd order polynomial fit based on fishtest data
// for two parameters needed to transform eval to the argument of a
// logistic function.
double as[] = {-8.24404295, 64.23892342, -95.73056462, 153.86478679};
double bs[] = {-3.37154371, 28.44489198, -56.67657741, 72.05858751};
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 eval to centipawns with limited range
double x = Utility::clamp(double(100 * v) / PawnValueEg, -1000.0, 1000.0);
// Return win rate in per mille
return 1000.0 / (1 + std::exp((a - x) / b));
}
// --------------------
// Call qsearch(),search() directly for testing
// --------------------