Version 1.16 comes with a significant strength increase over the last releasted version 1.15 and a lot less code.
Changes include:
-Better handling of endgames
-Retune parameters with Andrews data
-Late move reductions changes
-Add new evaluation features: King piece (shogi-like) incrementally updated tables.
Currently only for pawns.
Selfplay tests show:
1 thread:
ELO | 65.57 +- 4.74 (95%)
SPRT | 60.0+0.6s Threads=1 Hash=64MB
LLR | 0.00 (-2.94, 2.94) [0.00, 0.00]
Games | N: 12320 W: 4777 L: 2479 D: 5064
8 threads:
ELO | 74.03 +- 4.08 (95%)
SPRT | 20.0+0.2s Threads=8 Hash=256MB
LLR | 0.00 (-2.94, 2.94) [0.00, 0.00]
Games | N: 15050 W: 5625 L: 2466 D: 6959
Special thanks to
-Andrew Grant for his public dataset
-Terje (author of Weiss) for submitting several cleanup patches
-Connor (author of Seer) for the idea of using king piece tables
-Noob for really existing
This release wouldn't have been possible without them
If someone could test out the Linux compiles and check if they are working and that BMI2 faster POPCNT faster old on a modern cpu that would be nice :)