CasanChess 2.0.0 - new version chess engine
Casanchess - UCI compatible chess engine written in C++.
The NNUE evaluation is trained on millions of positions from auto-games and randomly generated FENs, evaluated with its own classical evaluation upto depth 7.
Author: Carlos Sanchez Mayordomo Rating CEDR=3029
CasanChess what's new?
A much stronger NNUE.
I spent the last month (++) improving the whole process: data generation, binary format, trainer, architecture... resulting in +230 self-Elo gain only from the new NNUE.
Moreover, now the evals are normalized (+1.0 eval corresponds to 75% expected score), and the endgame knowledge improves since the game results are used in the training, not only the evaluations.
Trained from 170M positions from self-played games at 250k nodes.
More info: #109
New UCI option: Ambition.
The NNUE also estimates the drawishness of the position. Casanchess uses this information to modify the search score and play less drawish continuations.
The default value 15 reduced draws from 48% to 41% in self-play but extra calculations cost around 3-5 Elo; the value 0 disables the feature and is slightly faster.
More info: #110
📦 The NNUE binary is now embedded in the executable.
Detailed changelog: #117
Diff: v1.1.2...v2.0
v2.0 vs v1.1
Elo | 256.41 +- 7.38 (95%)
Conf | 7.0+0.07s Threads=1 Hash=16MB
Games | N: 8004 W: 5486 L: 460 D: 2058
Penta | [12, 98, 596, 1444, 1852]
Elo | 228.34 +- 9.70 (95%)
Conf | 30.0+0.30s Threads=1 Hash=64MB
Games | N: 4000 W: 2542 L: 236 D: 1222
Penta | [7, 61, 345, 793, 794]
Comments
Post a Comment