Introduce static evaluation correction history

Vizvezdenec · Disservin · commit b4d995d0d910 · 2023-12-31T20:00:06.000+01:00
Idea from Caissa (https://github.com/Witek902/Caissa) chess engine. With given pawn structure collect data with how often search result and by how much it was better / worse than static evalution of position and use it to adjust static evaluation of positions with given pawn structure. Details: 1. excludes positions with fail highs and moves producing it being a capture; 2. update value is function of not only difference between best value and static evaluation but also is multiplied by linear function of depth; 3. maximum update value is maximum value of correction history divided by 2; 4. correction history itself is divided by 32 when applied so maximum value of static evaluation adjustment is 32 internal units. Passed STC: https://tests.stockfishchess.org/tests/view/658fc7b679aa8af82b955cac LLR: 2.96 (-2.94,2.94) <0.00,2.00> Total: 128672 W: 32757 L: 32299 D: 63616 Ptnml(0-2): 441, 15241, 32543, 15641, 470 Passed LTC: https://tests.stockfishchess.org/tests/view/65903f6979aa8af82b9566f1 LLR: 2.95 (-2.94,2.94) <0.50,2.50> Total: 97422 W: 24626 L: 24178 D: 48618 Ptnml(0-2): 41, 10837, 26527, 11245, 61 closes #4950 Bench: 1157852
diff --git a/src/movepick.cpp b/src/movepick.cpp
@@ -179,7 +179,7 @@ void MovePicker::score() {
 
             // histories
             m.value = 2 * (*mainHistory)[pos.side_to_move()][from_to(m)];
-            m.value += 2 * (*pawnHistory)[pawn_structure(pos)][pc][to];
+            m.value += 2 * (*pawnHistory)[pawn_structure_index(pos)][pc][to];
             m.value += 2 * (*continuationHistory[0])[pc][to];
             m.value += (*continuationHistory[1])[pc][to];
             m.value += (*continuationHistory[2])[pc][to] / 4;
@@ -216,7 +216,7 @@ void MovePicker::score() {
             else
                 m.value = (*mainHistory)[pos.side_to_move()][from_to(m)]
                         + (*continuationHistory[0])[pos.moved_piece(m)][to_sq(m)]
-                        + (*pawnHistory)[pawn_structure(pos)][pos.moved_piece(m)][to_sq(m)];
+                        + (*pawnHistory)[pawn_structure_index(pos)][pos.moved_piece(m)][to_sq(m)];
         }
 }
 
diff --git a/src/movepick.h b/src/movepick.h
@@ -33,12 +33,25 @@
 
 namespace Stockfish {
 
-constexpr int PAWN_HISTORY_SIZE = 512;  // has to be a power of 2
+constexpr int PAWN_HISTORY_SIZE        = 512;    // has to be a power of 2
+constexpr int CORRECTION_HISTORY_SIZE  = 16384;  // has to be a power of 2
+constexpr int CORRECTION_HISTORY_LIMIT = 1024;
 
 static_assert((PAWN_HISTORY_SIZE & (PAWN_HISTORY_SIZE - 1)) == 0,
               "PAWN_HISTORY_SIZE has to be a power of 2");
 
-inline int pawn_structure(const Position& pos) { return pos.pawn_key() & (PAWN_HISTORY_SIZE - 1); }
+static_assert((CORRECTION_HISTORY_SIZE & (CORRECTION_HISTORY_SIZE - 1)) == 0,
+              "CORRECTION_HISTORY_SIZE has to be a power of 2");
+
+enum PawnHistoryType {
+    Normal,
+    Correction
+};
+
+template<PawnHistoryType T = Normal>
+inline int pawn_structure_index(const Position& pos) {
+    return pos.pawn_key() & ((T == Normal ? PAWN_HISTORY_SIZE : CORRECTION_HISTORY_SIZE) - 1);
+}
 
 // StatsEntry stores the stat table value. It is usually a number but could
 // be a move or even a nested history. We use a class instead of a naked value
@@ -122,6 +135,10 @@ using ContinuationHistory = Stats<PieceToHistory, NOT_USED, PIECE_NB, SQUARE_NB>
 // PawnHistory is addressed by the pawn structure and a move's [piece][to]
 using PawnHistory = Stats<int16_t, 8192, PAWN_HISTORY_SIZE, PIECE_NB, SQUARE_NB>;
 
+// CorrectionHistory is addressed by color and pawn structure
+using CorrectionHistory =
+  Stats<int16_t, CORRECTION_HISTORY_LIMIT, COLOR_NB, CORRECTION_HISTORY_SIZE>;
+
 // MovePicker class is used to pick one pseudo-legal move at a time from the
 // current position. The most important method is next_move(), which returns a
 // new pseudo-legal move each time it is called, until there are no moves left,
diff --git a/src/search.cpp b/src/search.cpp
@@ -93,6 +93,11 @@ constexpr int futility_move_count(bool improving, Depth depth) {
     return improving ? (3 + depth * depth) : (3 + depth * depth) / 2;
 }
 
+// Guarantee evaluation does not hit the tablebase range
+constexpr Value to_static_eval(const Value v) {
+    return std::clamp(v, VALUE_TB_LOSS_IN_MAX_PLY + 1, VALUE_TB_WIN_IN_MAX_PLY - 1);
+}
+
 // History and stats update bonus, based on depth
 int stat_bonus(Depth d) { return std::min(268 * d - 352, 1153); }
 
@@ -712,6 +717,8 @@ Value search(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth, boo
 
     CapturePieceToHistory& captureHistory = thisThread->captureHistory;
 
+    Value unadjustedStaticEval = VALUE_NONE;
+
     // Step 6. Static evaluation of the position
     if (ss->inCheck)
     {
@@ -725,26 +732,40 @@ Value search(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth, boo
         // Providing the hint that this node's accumulator will be used often
         // brings significant Elo gain (~13 Elo).
         Eval::NNUE::hint_common_parent_position(pos);
-        eval = ss->staticEval;
+        unadjustedStaticEval = eval = ss->staticEval;
     }
     else if (ss->ttHit)
     {
         // Never assume anything about values stored in TT
-        ss->staticEval = eval = tte->eval();
+        unadjustedStaticEval = ss->staticEval = eval = tte->eval();
         if (eval == VALUE_NONE)
-            ss->staticEval = eval = evaluate(pos);
+            unadjustedStaticEval = ss->staticEval = eval = evaluate(pos);
         else if (PvNode)
             Eval::NNUE::hint_common_parent_position(pos);
 
+        Value newEval =
+          ss->staticEval
+          + thisThread->correctionHistory[us][pawn_structure_index<Correction>(pos)] / 32;
+
+        ss->staticEval = eval = to_static_eval(newEval);
+
         // ttValue can be used as a better position evaluation (~7 Elo)
         if (ttValue != VALUE_NONE && (tte->bound() & (ttValue > eval ? BOUND_LOWER : BOUND_UPPER)))
             eval = ttValue;
     }
     else
     {
-        ss->staticEval = eval = evaluate(pos);
-        // Save static evaluation into the transposition table
-        tte->save(posKey, VALUE_NONE, ss->ttPv, BOUND_NONE, DEPTH_NONE, MOVE_NONE, eval);
+        unadjustedStaticEval = ss->staticEval = eval = evaluate(pos);
+
+        Value newEval =
+          ss->staticEval
+          + thisThread->correctionHistory[us][pawn_structure_index<Correction>(pos)] / 32;
+
+        ss->staticEval = eval = to_static_eval(newEval);
+
+        // Static evaluation is saved as it was before adjustment by correction history
+        tte->save(posKey, VALUE_NONE, ss->ttPv, BOUND_NONE, DEPTH_NONE, MOVE_NONE,
+                  unadjustedStaticEval);
     }
 
     // Use static evaluation difference to improve quiet move ordering (~9 Elo)
@@ -753,7 +774,8 @@ Value search(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth, boo
         int bonus = std::clamp(-13 * int((ss - 1)->staticEval + ss->staticEval), -1652, 1546);
         thisThread->mainHistory[~us][from_to((ss - 1)->currentMove)] << bonus;
         if (type_of(pos.piece_on(prevSq)) != PAWN && type_of((ss - 1)->currentMove) != PROMOTION)
-            thisThread->pawnHistory[pawn_structure(pos)][pos.piece_on(prevSq)][prevSq] << bonus / 4;
+            thisThread->pawnHistory[pawn_structure_index(pos)][pos.piece_on(prevSq)][prevSq]
+              << bonus / 4;
     }
 
     // Set up the improving flag, which is true if current static evaluation is
@@ -888,7 +910,7 @@ Value search(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth, boo
                 {
                     // Save ProbCut data into transposition table
                     tte->save(posKey, value_to_tt(value, ss->ply), ss->ttPv, BOUND_LOWER, depth - 3,
-                              move, ss->staticEval);
+                              move, unadjustedStaticEval);
                     return std::abs(value) < VALUE_TB_WIN_IN_MAX_PLY ? value - (probCutBeta - beta)
                                                                      : value;
                 }
@@ -999,10 +1021,10 @@ Value search(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth, boo
             }
             else
             {
-                int history = (*contHist[0])[movedPiece][to_sq(move)]
-                            + (*contHist[1])[movedPiece][to_sq(move)]
-                            + (*contHist[3])[movedPiece][to_sq(move)]
-                            + thisThread->pawnHistory[pawn_structure(pos)][movedPiece][to_sq(move)];
+                int history =
+                  (*contHist[0])[movedPiece][to_sq(move)] + (*contHist[1])[movedPiece][to_sq(move)]
+                  + (*contHist[3])[movedPiece][to_sq(move)]
+                  + thisThread->pawnHistory[pawn_structure_index(pos)][movedPiece][to_sq(move)];
 
                 // Continuation history based pruning (~2 Elo)
                 if (lmrDepth < 6 && history < -3752 * depth)
@@ -1364,12 +1386,23 @@ Value search(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth, boo
         ss->ttPv = ss->ttPv || ((ss - 1)->ttPv && depth > 3);
 
     // Write gathered information in transposition table
+    // Static evaluation is saved as it was before correction history
     if (!excludedMove && !(rootNode && thisThread->pvIdx))
         tte->save(posKey, value_to_tt(bestValue, ss->ply), ss->ttPv,
                   bestValue >= beta    ? BOUND_LOWER
                   : PvNode && bestMove ? BOUND_EXACT
                                        : BOUND_UPPER,
-                  depth, bestMove, ss->staticEval);
+                  depth, bestMove, unadjustedStaticEval);
+
+    // Adjust correction history
+    if (!ss->inCheck && (!bestMove || !pos.capture(bestMove))
+        && !(bestValue >= beta && bestValue <= ss->staticEval)
+        && !(!bestMove && bestValue >= ss->staticEval))
+    {
+        auto bonus = std::clamp(int(bestValue - ss->staticEval) * depth / 8,
+                                -CORRECTION_HISTORY_LIMIT / 4, CORRECTION_HISTORY_LIMIT / 4);
+        thisThread->correctionHistory[us][pawn_structure_index<Correction>(pos)] << bonus;
+    }
 
     assert(bestValue > -VALUE_INFINITE && bestValue < VALUE_INFINITE);
 
@@ -1450,6 +1483,8 @@ Value qsearch(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth) {
         && (tte->bound() & (ttValue >= beta ? BOUND_LOWER : BOUND_UPPER)))
         return ttValue;
 
+    Value unadjustedStaticEval = VALUE_NONE;
+
     // Step 4. Static evaluation of the position
     if (ss->inCheck)
         bestValue = futilityBase = -VALUE_INFINITE;
@@ -1458,25 +1493,39 @@ Value qsearch(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth) {
         if (ss->ttHit)
         {
             // Never assume anything about values stored in TT
-            if ((ss->staticEval = bestValue = tte->eval()) == VALUE_NONE)
-                ss->staticEval = bestValue = evaluate(pos);
+            if ((unadjustedStaticEval = ss->staticEval = bestValue = tte->eval()) == VALUE_NONE)
+                unadjustedStaticEval = ss->staticEval = bestValue = evaluate(pos);
+
+            Value newEval =
+              ss->staticEval
+              + thisThread->correctionHistory[us][pawn_structure_index<Correction>(pos)] / 32;
+
+            ss->staticEval = bestValue = to_static_eval(newEval);
 
             // ttValue can be used as a better position evaluation (~13 Elo)
             if (ttValue != VALUE_NONE
                 && (tte->bound() & (ttValue > bestValue ? BOUND_LOWER : BOUND_UPPER)))
                 bestValue = ttValue;
         }
         else
+        {
             // In case of null move search, use previous static eval with a different sign
-            ss->staticEval = bestValue =
+            unadjustedStaticEval = ss->staticEval = bestValue =
               (ss - 1)->currentMove != MOVE_NULL ? evaluate(pos) : -(ss - 1)->staticEval;
 
+            Value newEval =
+              ss->staticEval
+              + thisThread->correctionHistory[us][pawn_structure_index<Correction>(pos)] / 32;
+
+            ss->staticEval = bestValue = to_static_eval(newEval);
+        }
+
         // Stand pat. Return immediately if static value is at least beta
         if (bestValue >= beta)
         {
             if (!ss->ttHit)
                 tte->save(posKey, value_to_tt(bestValue, ss->ply), false, BOUND_LOWER, DEPTH_NONE,
-                          MOVE_NONE, ss->staticEval);
+                          MOVE_NONE, unadjustedStaticEval);
 
             return bestValue;
         }
@@ -1620,8 +1669,10 @@ Value qsearch(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth) {
         bestValue = (3 * bestValue + beta) / 4;
 
     // Save gathered info in transposition table
+    // Static evaluation is saved as it was before adjustment by correction history
     tte->save(posKey, value_to_tt(bestValue, ss->ply), pvHit,
-              bestValue >= beta ? BOUND_LOWER : BOUND_UPPER, ttDepth, bestMove, ss->staticEval);
+              bestValue >= beta ? BOUND_LOWER : BOUND_UPPER, ttDepth, bestMove,
+              unadjustedStaticEval);
 
     assert(bestValue > -VALUE_INFINITE && bestValue < VALUE_INFINITE);
 
@@ -1720,15 +1771,17 @@ void update_all_stats(const Position& pos,
 
         // Increase stats for the best move in case it was a quiet move
         update_quiet_stats(pos, ss, bestMove, bestMoveBonus);
-        thisThread->pawnHistory[pawn_structure(pos)][moved_piece][to_sq(bestMove)]
-          << quietMoveBonus;
+
+        int pIndex = pawn_structure_index(pos);
+        thisThread->pawnHistory[pIndex][moved_piece][to_sq(bestMove)] << quietMoveBonus;
 
         // Decrease stats for all non-best quiet moves
         for (int i = 0; i < quietCount; ++i)
         {
-            thisThread->pawnHistory[pawn_structure(pos)][pos.moved_piece(quietsSearched[i])]
-                                   [to_sq(quietsSearched[i])]
+            thisThread
+                ->pawnHistory[pIndex][pos.moved_piece(quietsSearched[i])][to_sq(quietsSearched[i])]
               << -quietMoveMalus;
+
             thisThread->mainHistory[us][from_to(quietsSearched[i])] << -quietMoveMalus;
             update_continuation_histories(ss, pos.moved_piece(quietsSearched[i]),
                                           to_sq(quietsSearched[i]), -quietMoveMalus);
diff --git a/src/thread.cpp b/src/thread.cpp
@@ -70,6 +70,7 @@ void Thread::clear() {
     mainHistory.fill(0);
     captureHistory.fill(0);
     pawnHistory.fill(0);
+    correctionHistory.fill(0);
 
     for (bool inCheck : {false, true})
         for (StatsType c : {NoCaptures, Captures})
diff --git a/src/thread.h b/src/thread.h
@@ -69,6 +69,7 @@ class Thread {
     CapturePieceToHistory captureHistory;
     ContinuationHistory   continuationHistory[2][2];
     PawnHistory           pawnHistory;
+    CorrectionHistory     correctionHistory;
 };
 
 
