Tuned history reduction
authorLyudmil Antonov <lantonov.here@gmail.com>
Mon, 17 Feb 2020 09:13:03 +0000 (11:13 +0200)
committerJoost VandeVondele <Joost.VandeVondele@gmail.com>
Thu, 26 Mar 2020 16:49:28 +0000 (17:49 +0100)
Value after a long Bayesian tuning, using a home-made classification approach.

STC
https://tests.stockfishchess.org/tests/view/5e7c7b16e42a5c3b3ca2eb66
LLR: 2.96 (-2.94,2.94) {-0.50,1.50}
Total: 45472 W: 8992 L: 8732 D: 27748
Ptnml(0-2): 795, 5276, 10352, 5500, 813

LTC
https://tests.stockfishchess.org/tests/view/5e7c8be7e42a5c3b3ca2eb75
LLR: 2.94 (-2.94,2.94) {0.25,1.75}
Total: 22744 W: 3085 L: 2849 D: 16810
Ptnml(0-2): 156, 2090, 6658, 2298, 170

closes https://github.com/official-stockfish/Stockfish/pull/2597

Bench 5030855

src/search.cpp

index abc6874e99583360a679ffe1cdad8806fee34861..c5a7582ce4d1ccea45a67804983c3fde992bc3dc 100644 (file)
@@ -1182,7 +1182,7 @@ moves_loop: // When in check, search starts from here
                   r++;
 
               // Decrease/increase reduction for moves with a good/bad history (~30 Elo)
-              r -= ss->statScore / 16384;
+              r -= ss->statScore / 16434;
           }
 
           // Increase reduction for captures/promotions if late move and at low depth