Update neural network trainer and bump to version 1.0

README.md

@@ -3,11 +3,9 @@ UCI command-line chess engine written in C++ from scratch. In development since
 
 ### Playing strength
 
-Estimated score of the newest version: **~3290 Elo**
+CCRL 40/15 Score: **3293** (#44) (version 0.9.0)
 
-CCRL 40/15 Score: **3218** (#45) (version 0.8.0)
-
-CCRL 2+1 Score: **3191** (#62) (version 0.7.0)
+CCRL 2+1 Score: **3222** (#46) (version 0.8.0)
 
 ### Supported UCI options
 

src/backend/Evaluate.cpp

@@ -12,8 +12,8 @@
 #include <fstream>
 #include <memory>
 
-const char* c_DefaultEvalFile = "eval-4.pnn";
-const char* c_DefaultEndgameEvalFile = "endgame-3.pnn";
+const char* c_DefaultEvalFile = "eval.pnn";
+const char* c_DefaultEndgameEvalFile = "endgame.pnn";
 
 #define S(mg, eg) PieceScore{ mg, eg }
 

src/frontend/UCI.cpp

@@ -10,7 +10,7 @@
 #include <math.h>
 #include <random>
 
-#define VersionNumber "0.9.5"
+#define VersionNumber "1.0"
 
 #if defined(USE_BMI2) && defined(USE_AVX2) 
 #define ArchitectureStr "AVX2/BMI2"

src/frontend/frontend.rc

@@ -25,8 +25,8 @@ LANGUAGE LANG_POLISH, SUBLANG_DEFAULT
 //
 
 VS_VERSION_INFO VERSIONINFO
- FILEVERSION 0,4,0,0
- PRODUCTVERSION 0,4,0,0
+ FILEVERSION 1,0,0,0
+ PRODUCTVERSION 1,0,0,0
  FILEFLAGSMASK 0x3fL
 #ifdef _DEBUG
  FILEFLAGS 0x3L
@@ -42,12 +42,12 @@ BEGIN
         BLOCK "000904b0"
         BEGIN
             VALUE "FileDescription", "Caissa Chess Engine"
-            VALUE "FileVersion", "0.4.0.0"
+            VALUE "FileVersion", "1.0.0.0"
             VALUE "InternalName", "frontend.exe"
-            VALUE "LegalCopyright", "Copyright (C) 2021"
+            VALUE "LegalCopyright", "Copyright (C) 2022"
             VALUE "OriginalFilename", "Caissa.exe"
             VALUE "ProductName", "Caissa"
-            VALUE "ProductVersion", "0.4.0.0"
+            VALUE "ProductVersion", "1.0.0.0"
         END
     END
     BLOCK "VarFileInfo"

src/utils/NetworkTrainer.cpp

@@ -79,6 +79,8 @@ bool TrainNetwork()
         networksData[i].runCtx.Init(networksData[i].network);
     }
 
+    std::ofstream trainingLog("training.log");
+
     std::vector<PositionEntry> entries;
     LoadAllPositions(entries);
 
@@ -278,6 +280,8 @@ bool TrainNetwork()
                 networksData[i].packedNet.Save((name + ".pnn").c_str());
                 networksData[i].packedNet.SaveAsImage((name + ".raw").c_str());
             }
+
+            trainingLog << iteration << "\t" << nnErrorSum << "\t" << nnPackedErrorSum << std::endl;
         }
     }
 

src/utils/NeuralNetwork.cpp

@@ -633,11 +633,11 @@ void Layer::UpdateWeights_AdaDelta(float learningRate, const Gradients& gradient
 
 void NeuralNetwork::ClampLayerWeights(size_t layerIndex, float weightRange, float biasRange, float weightQuantizationScale, float biasQuantizationScale)
 {
-    const float cDecay = 1.0e-7f;
+    const float cDecay = 0.5e-6f;
 
     Layer& layer = layers[layerIndex];
 
-    for (uint32_t j = 0; j < layer.numInputs; j++)
+    for (uint32_t j = 0; j <= layer.numInputs; j++)
     {
         const bool isBiasWeight = (j == layer.numInputs);
 
@@ -649,11 +649,11 @@ void NeuralNetwork::ClampLayerWeights(size_t layerIndex, float weightRange, floa
 
             if (isBiasWeight)
             {
-                w = std::clamp(w * biasQuantizationScale, -biasRange, biasRange) / biasQuantizationScale;
+                w = std::clamp(w, -biasRange / biasQuantizationScale, biasRange / biasQuantizationScale) ;
             }
             else
             {
-                w = std::clamp(w * weightQuantizationScale, -weightRange, weightRange) / weightQuantizationScale;
+                w = std::clamp(w, -weightRange / weightQuantizationScale, weightRange / weightQuantizationScale) ;
             }
         }
     }
@@ -731,7 +731,7 @@ void NeuralNetworkTrainer::Train(NeuralNetwork& network, const TrainingSet& trai
                     for (size_t i = 0; i < ctx.tempValues.size(); i++)
                     {
                         // gradient of RMS loss function
-                        ctx.tempValues[i] = ctx.tempValues[i] - vec.output[i];
+                        ctx.tempValues[i] = /* 2.0f * */ (ctx.tempValues[i] - vec.output[i]);
 
                         // gradient of cross-entropy loss function
                         //const float target = vec.output[i];

src/utils/TrainerCommon.cpp

@@ -59,7 +59,7 @@ static bool LoadPositions(const char* fileName, std::vector<PositionEntry>& entr
 
             if (move.IsQuiet() &&
                 pos.GetNumPieces() >= 6 &&
-                pos.GetHalfMoveCount() < 80 &&
+                pos.GetHalfMoveCount() < 60 &&
                 whitePawnsMoved && blackPawnsMoved &&
                 !pos.IsInCheck() && pos.GetNumLegalMoves())
             {
@@ -77,7 +77,7 @@ static bool LoadPositions(const char* fileName, std::vector<PositionEntry>& entr
                     // blend in future scores into current move score
                     float scoreSum = 0.0f;
                     float weightSum = 0.0f;
-                    const size_t maxLookahead = 16;
+                    const size_t maxLookahead = 12;
                     for (size_t j = 0; j < maxLookahead; ++j)
                     {
                         if (i + j >= game.GetMoves().size()) break;
@@ -99,6 +99,9 @@ static bool LoadPositions(const char* fileName, std::vector<PositionEntry>& entr
                     entry.score = std::lerp(score, scoreSum, lambda);
                 }
 
+                const float offset = 0.00001f;
+                entry.score = offset + entry.score * (1.0f - 2.0f * offset);
+
                 Position normalizedPos = pos;
                 if (pos.GetSideToMove() == Color::Black)
                 {