Improving flag via GrandParent nodes. Adjust LMR to use a new formula + legal moves. (#413)

Bench: 4467773

STC

ELO   | 4.05 +- 3.18 (95%)
SPRT  | 10.0+0.10s Threads=1 Hash=8MB
LLR   | 3.00 (-2.94, 2.94) [-0.50, 3.00]
GAMES | N: 21336 W: 5099 L: 4850 D: 11387

LTC

ELO   | 3.37 +- 2.49 (95%)
SPRT  | 60.0+0.60s Threads=1 Hash=64MB
LLR   | 2.98 (-2.94, 2.94) [0.00, 3.00]
GAMES | N: 32536 W: 7232 L: 6916 D: 18388

Author: jhonnold

src/makefile

@@ -4,7 +4,7 @@
 EXE      = berserk
 SRC      = *.c pyrrhic/tbprobe.c
 CC       = gcc
-VERSION  = 20221001
+VERSION  = 20221003
 MAIN_NETWORK = networks/berserk-c982d9682d4e.nn
 EVALFILE = $(MAIN_NETWORK)
 DEFS     = -DVERSION=\"$(VERSION)\" -DEVALFILE=\"$(EVALFILE)\" -DNDEBUG

src/search.c

@@ -49,7 +49,7 @@ extern volatile int PONDERING;
 
 void InitPruningAndReductionTables() {
   for (int depth = 1; depth < MAX_SEARCH_PLY; depth++)
-    for (int moves = 1; moves < 64; moves++) LMR[depth][moves] = log(depth) * log(moves) / 2 - 0.2;
+    for (int moves = 1; moves < 64; moves++) LMR[depth][moves] = log(depth) * log(moves) / 2.25 + 0.25;
 
   LMR[0][0] = LMR[0][1] = LMR[1][0] = 0;
 
@@ -389,8 +389,14 @@ int Negamax(int alpha, int beta, int depth, int cutnode, ThreadData* thread, PV*
   if (!tt) TTPut(board->zobrist, INT8_MIN, UNKNOWN, TT_UNKNOWN, NULL_MOVE, data->ply, eval, ttPv);
 
   // getting better if eval has gone up
-  int improving = !board->checkers && data->ply >= 2 &&
-                  (data->evals[data->ply] > data->evals[data->ply - 2] || data->evals[data->ply - 2] == UNKNOWN);
+  int improving = 0;
+  if (!board->checkers && data->ply >= 2) {
+    if (data->ply >= 4 && data->evals[data->ply - 2] == UNKNOWN) {
+      improving = data->evals[data->ply] > data->evals[data->ply - 4] || data->evals[data->ply - 4] == UNKNOWN;
+    } else {
+      improving = data->evals[data->ply] > data->evals[data->ply - 2] || data->evals[data->ply - 2] == UNKNOWN;
+    }
+  }
 
   // reset moves to moves related to 1 additional ply
   data->skipMove[data->ply + 1]   = NULL_MOVE;
@@ -560,8 +566,8 @@ int Negamax(int alpha, int beta, int depth, int cutnode, ThreadData* thread, PV*
 
     // Late move reductions
     int R = 1;
-    if (depth > 2 && playedMoves > 1 && !tactical) {
-      R = LMR[min(depth, 63)][min(playedMoves, 63)];
+    if (depth > 2 && legalMoves > 1 && !tactical) {
+      R = LMR[min(depth, 63)][min(legalMoves, 63)];
 
       // increase reduction on non-pv
       if (!ttPv) R++;