Update the Network, trained differently and on FRC data

The previous network was trained to output a score, which was, in training, applied to both the MG and EG components of the evaluation. However, in implementation, we were only applying it to the MG in Ethereal.

This new network is trained to expressly produce a score only for the MG, and to have it only be applied to the MG component of the eval. This is done correctly.

Additionally, the Network is trained using an additionally ~9.6 million FRC positions. This was done to increase diversity, and to resolve an issue where Regression Tests failed to show FRC improvements for the previous network.

Testing was done to confirm that the gains for FRC are massive. Testing was done to confirm that any changes to Standard chess are minimal.

Tests for Fischer Random Chess:

ELO   | 18.90 +- 9.10 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=8MB
LLR   | 2.97 (-2.94, 2.94) [0.00, 5.00]
Games | N: 2816 W: 785 L: 632 D: 1399
http://chess.grantnet.us/test/7548/

ELO   | 18.95 +- 8.32 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 3.04 (-2.94, 2.94) [0.00, 5.00]
Games | N: 2624 W: 586 L: 443 D: 1595
http://chess.grantnet.us/test/7549/

Tests for Standard Chess:

ELO   | 1.14 +- 2.53 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=8MB
LLR   | -0.93 (-2.94, 2.94) [0.00, 5.00]
Games | N: 29376 W: 6003 L: 5907 D: 17466
http://chess.grantnet.us/test/7547/

ELO   | 1.63 +- 2.27 (95%)
SPRT  | 10.0+0.1s Threads=1 Hash=8MB
LLR   | 0.25 (-2.94, 2.94) [0.00, 5.00]
Games | N: 33296 W: 6281 L: 6125 D: 20890
http://chess.grantnet.us/test/7546/

BENCH : 4,933,831

src/network.c

@@ -24,6 +24,7 @@
 
 #include "bitboards.h"
 #include "board.h"
+#include "evaluate.h"
 #include "move.h"
 #include "network.h"
 #include "thread.h"
@@ -112,11 +113,13 @@ int fullyComputePKNetwork(Thread *thread) {
                 outputNeurons[i] += layer1Neurons[j] * PKNN.layer1Weights[i][j];
     }
 
-    return outputNeurons[0];
+    return MakeScore((int) outputNeurons[0], 0);
 }
 
 int partiallyComputePKNetwork(Thread *thread) {
 
+    if (TRACE) return fullyComputePKNetwork(thread);
+
     float *layer1Neurons = thread->pknnlayer1[thread->pknndepth];
     float outputNeurons[PKNETWORK_OUTPUTS];
 
@@ -127,7 +130,7 @@ int partiallyComputePKNetwork(Thread *thread) {
                 outputNeurons[i] += layer1Neurons[j] * PKNN.layer1Weights[i][j];
     }
 
-    return outputNeurons[0];
+    return MakeScore((int) outputNeurons[0], 0);
 }
 
 
@@ -183,7 +186,6 @@ void updatePKNetworkAfterMove(Thread *thread, uint16_t move) {
         }
 
         if (taken == PAWN) {
-            assert(0);
             indexes[changes++] = computePKNetworkIndex(!colour, taken, to);
             signs[changes - 1] = -1;
         }

src/network.h

@@ -31,9 +31,7 @@ typedef struct PKNetwork {
 
     // PKNetworks are of the form [Input, Hidden Layer 1, Output Layer]
     // Our current Network is [224x32, 32x1]. The Network is trained to
-    // output a Score in CentiPawns, and thus Output Neurons need do not
-    // need activation functions. The 32x1 operation applys a ReLU, where
-    // as the 224x32 layer does not, since inputs are binary 0s and 1s
+    // output a Score in CentiPawns for the Midgame only.
 
     ALIGN64 float inputWeights[PKNETWORK_LAYER1][PKNETWORK_INPUTS];
     ALIGN64 float inputBiases[PKNETWORK_LAYER1];

src/tuner.c

@@ -205,12 +205,9 @@ void initCoefficients(TVector coeffs) {
 
 void initTunerEntries(TEntry *entries, Thread *thread, TArray methods) {
 
-    Undo undo;
     char line[256];
-    Limits limits = {0};
-    thread->limits = &limits; thread->depth  = 0;
-
     FILE *fin = fopen("FENS", "r");
+
     for (int i = 0; i < NPOSITIONS; i++) {
 
         if (fgets(line, 256, fin) == NULL)
@@ -225,14 +222,7 @@ void initTunerEntries(TEntry *entries, Thread *thread, TArray methods) {
         // Set the board with the current FEN
         boardFromFEN(&thread->board, line, 0);
 
-        // Resolve the position to mitigate tactics
-        if (QSRESOLVE) {
-            qsearch(thread, &thread->pv, -MATE, MATE, 0);
-            for (int pvidx = 0; pvidx < thread->pv.length; pvidx++)
-                applyMove(&thread->board, thread->pv.line[pvidx], &undo);
-        }
-
-        // Defer the set to another function
+        // Defer the setup to another function
         initTunerEntry(&entries[i], thread, &thread->board, methods);
 
         // Occasional reporting for total completion
@@ -324,13 +314,15 @@ double computeOptimalK(TEntry *entries) {
                 best = error, start = curr;
         }
 
-        printf("Epoch [%d] K = %f E = %f\n", i, start, best);
+        printf("Epoch [%d] K = [%.9f] E = [%.9f]\n", i, start, best);
 
         end   = start + step;
         start = start - step;
         step  = step  / 10.0;
     }
 
+    printf("\n");
+
     return start;
 }
 

src/tuner.h

@@ -22,25 +22,23 @@
 
 #include "types.h"
 
-#define NPARTITIONS     (      64) // Total thread partitions
-#define KPRECISION      (      10) // Iterations for computing K
-
-#define QSRESOLVE       (       0) // Whether to resolve via a qsearch()
-#define PRETTYIFY       (       1) // Whether to format as if we tune everything
-#define REPORTING       (      50) // How often to print the new parameters
-
-#define LRRATE          (    0.10) // Global Learning rate
-#define LRDROPRATE      (    1.00) // Cut LR by this each LR-step
-#define LRSTEPRATE      (     250) // Cut LR after this many epochs
-
-#define TuneNormal     (        0) // Flag to enable tuning on all Normals
-#define TuneSafety     (        0) // Flag to enable tuning on all Safeties
-#define TuneComplexity (        0) // Flag to enable tuning on all Complexities
-
-#define NTERMS         (       0) // Total terms in the Tuner (904)
-#define MAXEPOCHS      (   10000) // Max number of epochs allowed
-#define BATCHSIZE      (   16384) // FENs per mini-batch
-#define NPOSITIONS     ( 9999740) // Total FENS in the book
+#define NPARTITIONS    (      64) // Total thread partitions
+#define KPRECISION     (      10) // Iterations for computing K
+#define PRETTYIFY      (       0) // Whether to format as if we tune everything
+#define REPORTING      (      50) // How often to print the new parameters
+
+#define LRRATE         (    0.10) // Global Learning rate
+#define LRDROPRATE     (    1.00) // Cut LR by this each LR-step
+#define LRSTEPRATE     (     250) // Cut LR after this many epochs
+
+#define TuneNormal     (       0) // Flag to enable all Normals      (856)
+#define TuneSafety     (       0) // Flag to enable all Safeties     ( 44)
+#define TuneComplexity (       0) // Flag to enable all Complexities (  4)
+
+#define NTERMS         (      0) // Total terms in the Tuner (904)
+#define MAXEPOCHS      (  100000) // Max number of epochs allowed
+#define BATCHSIZE      (   16384) // Training samples per mini-batch
+#define NPOSITIONS     ( 9999740) // Total Training samples in the book
 
 #define STACKSIZE ((int)((double) NPOSITIONS * NTERMS / 64))
 

src/uci.h

@@ -22,7 +22,7 @@
 
 #include "types.h"
 
-#define VERSION_ID "12.60"
+#define VERSION_ID "12.61"
 
 #if defined(USE_PEXT)
     #define ETHEREAL_VERSION VERSION_ID" (PEXT)"