Revert "use winning_percentage_wdl in learn"

This reverts commit c964e902c5.

# Conflicts:
#	src/uci.cpp

src/learn/learner.cpp

@@ -116,7 +116,6 @@ bool use_draw_in_training_data_generation = false;
 bool use_draw_in_training = false;
 bool use_draw_in_validation = false;
 bool use_hash_in_training = true;
-bool use_wdl = false;
 
 // -----------------------------------
 // write phase file
@@ -1026,16 +1025,6 @@ double sigmoid(double x)
 	return 1.0 / (1.0 + std::exp(-x));
 }
 
-// A function that converts the evaluation value to the winning rate [0,1]
-double winning_percentage_wdl(Value value, int ply)
-{
-	double wdl_w = UCI::win_rate_model( value, ply);
-	double wdl_l = UCI::win_rate_model(-value, ply);
-	double wdl_d = 1000.0 - wdl_w - wdl_l;
-
-	return (wdl_w + wdl_d / 2.0) / 1000.0;
-}
-
 // A function that converts the evaluation value to the winning rate [0,1]
 double winning_percentage(double value)
 {
@@ -1044,18 +1033,6 @@ double winning_percentage(double value)
 	// = sigmoid(Eval/4*ln(10))
 	return sigmoid(value / PawnValueEg / 4.0 * log(10.0));
 }
-
-// A function that converts the evaluation value to the winning rate [0,1]
-double winning_percentage(Value value, int ply)
-{
-	if (use_wdl) {
-		return winning_percentage_wdl(value, ply);
-	}
-	else {
-		return winning_percentage(value);
-	}
-}
-
 double dsigmoid(double x)
 {
 	// Sigmoid function
@@ -1092,8 +1069,8 @@ double calc_grad(Value deep, Value shallow, PackedSfenValue& psv)
 	// Also, the coefficient of 1/m is unnecessary if you use the update formula that has the automatic gradient adjustment function like Adam and AdaGrad.
 	// Therefore, it is not necessary to save it in memory.
 
-	double p = winning_percentage(deep, psv.gamePly);
-	double q = winning_percentage(shallow, psv.gamePly);
+	double p = winning_percentage(deep);
+	double q = winning_percentage(shallow);
 	return (q - p) * dsigmoid(double(shallow) / 600.0);
 }
 #endif
@@ -1118,8 +1095,8 @@ double calc_grad(Value deep, Value shallow, const PackedSfenValue& psv)
 	// = ...
 	// = q-p.
 
-	double p = winning_percentage(deep, psv.gamePly);
-	double q = winning_percentage(shallow, psv.gamePly);
+	double p = winning_percentage(deep);
+	double q = winning_percentage(shallow);
 
 	return q - p;
 }
@@ -1150,8 +1127,8 @@ double calc_grad(Value deep, Value shallow , const PackedSfenValue& psv)
 	// elmo (WCSC27) method
 	// Correct with the actual game wins and losses.
 
-	const double q = winning_percentage(shallow, psv.gamePly);
-	const double p = winning_percentage(deep, psv.gamePly);
+	const double q = winning_percentage(shallow);
+	const double p = winning_percentage(deep);
 
 	// Use 1 as the correction term if the expected win rate is 1, 0 if you lose, and 0.5 if you draw.
 	// game_result = 1,0,-1 so add 1 and divide by 2.
@@ -1173,8 +1150,8 @@ void calc_cross_entropy(Value deep, Value shallow, const PackedSfenValue& psv,
 	double& cross_entropy_eval, double& cross_entropy_win, double& cross_entropy,
 	double& entropy_eval, double& entropy_win, double& entropy)
 {
-	const double p /* teacher_winrate */ = winning_percentage(deep, psv.gamePly);
-	const double q /* eval_winrate    */ = winning_percentage(shallow, psv.gamePly);
+	const double p /* teacher_winrate */ = winning_percentage(deep);
+	const double q /* eval_winrate    */ = winning_percentage(shallow);
 	const double t = double(psv.game_result + 1) / 2;
 
 	constexpr double epsilon = 0.000001;
@@ -2943,7 +2920,6 @@ void learn(Position&, istringstream& is)
 		else if (option == "use_draw_in_training") is >> use_draw_in_training;
 		else if (option == "use_draw_in_validation") is >> use_draw_in_validation;
 		else if (option == "use_hash_in_training") is >> use_hash_in_training;
-		else if (option == "use_wdl") is >> use_wdl;
 		// Discount rate
 		else if (option == "discount_rate") is >> discount_rate;
 

src/uci.cpp

@@ -248,32 +248,6 @@ namespace {
   uint64_t eval_sum;
 } // namespace
 
-
-namespace UCI{
-  // 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)));
-  }
-} // namespace UCI
-
-
 // Make is_ready_cmd() callable from outside. (Because I want to call it from the bench command etc.)
 // Note that the phase is not initialized.
 void is_ready(bool skipCorruptCheck)

src/uci.h

@@ -74,7 +74,6 @@ std::string square(Square s);
 std::string move(Move m, bool chess960);
 std::string pv(const Position& pos, Depth depth, Value alpha, Value beta);
 std::string wdl(Value v, int ply);
-int win_rate_model(Value v, int ply);
 Move to_move(const Position& pos, std::string& str);
 
 // Flag that read the evaluation function. This is set to false when evaldir is changed.