LMR Simplification
authorElbertoOne <ElbertoOne@users.noreply.github.com>
Fri, 3 Jun 2016 08:01:02 +0000 (10:01 +0200)
committerMarco Costalba <mcostalba@gmail.com>
Fri, 3 Jun 2016 17:53:04 +0000 (19:53 +0200)
LMR simplification that also gives a slight ELO gain, especially at LTC:

STC (http://tests.stockfishchess.org/tests/view/574ec8e20ebc59029919b147):
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 32402 W: 5967 L: 5866 D: 20569

LTC (http://tests.stockfishchess.org/tests/view/574fbebf0ebc59029919b16d):
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 15103 W: 2103 L: 1975 D: 11025

Bench: 8248133

src/search.cpp

index c6d63c7..5f0a0f3 100644 (file)
@@ -1019,19 +1019,18 @@ moves_loop: // When in check search starts from here
           if (!PvNode && cutNode)
               r += ONE_PLY;
 
-          // Decrease/increase reduction for moves with a good/bad history
-          int rHist = (val - 10000) / 20000;
-          r = std::max(DEPTH_ZERO, r - rHist * ONE_PLY);
-
           // Decrease reduction for moves that escape a capture. Filter out
           // castling moves, because they are coded as "king captures rook" and
           // hence break make_move(). Also use see() instead of see_sign(),
           // because the destination square is empty.
-          if (   r
-              && type_of(move) == NORMAL
-              && type_of(pos.piece_on(to_sq(move))) != PAWN
-              && pos.see(make_move(to_sq(move), from_sq(move))) < VALUE_ZERO)
-              r = std::max(DEPTH_ZERO, r - ONE_PLY);
+          else if (   type_of(move) == NORMAL
+                   && type_of(pos.piece_on(to_sq(move))) != PAWN
+                   && pos.see(make_move(to_sq(move), from_sq(move))) < VALUE_ZERO)
+              r -= ONE_PLY;
+
+          // Decrease/increase reduction for moves with a good/bad history
+          int rHist = (val - 10000) / 20000;
+          r = std::max(DEPTH_ZERO, r - rHist * ONE_PLY);
 
           Depth d = std::max(newDepth - r, ONE_PLY);