Obsidian, an UCI chess engine written in c++
Rating Chess Engines Diary CEDR=3707
v.14 SE:
Obsidian comes back far stronger than ever:
this release features a few bug fixes, multithreaded TT initialization, and especially, Elo gains I wasn't sure I would've been able to delivier.
Both tests below were run at a way higher time control than is written - I won't get into the details of this here - it doesn't matter.
Obsidian 14.0 vs 13.0 (STC)
Elo | 40.34 +- 6.59 (95%)
Conf | 8.0+0.08s Threads=1 Hash=16MB
Games | N: 3002 W: 947 L: 600 D: 1455
Penta | [6, 227, 712, 526, 30]
Obsidian 14.0 vs 13.0 (LTC)
Elo | 46.25 +- 6.05 (95%)
Conf | 40.0+0.40s Threads=1 Hash=128MB
Games | N: 3000 W: 899 L: 502 D: 1599
Penta | [1, 184, 740, 567, 8]
Obsidian 13 - results:
Opponent | Score | +- | Games |
PlentyChess 3.0.0 | 119.5/240 | -1 | 240 Partii |
Alexandria 7.0.0 | 58.5/109 | +8 | 109 Partii |
Chess-System-Tal-2 | 55.5/97 | +14 | 97 Partii |
Berserk 13 | 45.5/95 | -4 | 95 Partii |
Ethereal 14.25 | 45.5/83 | +8 | 83 Partii |
Titan 1.1.0 | 39.5/68 | +11 | 68 Partii |
RubiChess 20240112 | 34.5/65 | +4 | 65 Partii |
Uralochka 3.41a JA | 38/64 | +12 | 64 Partii |
Lizard 10.5 | 36.5/64 | +9 | 64 Partii |
Caissa 1.19 | 34/63 | +5 | 63 Partii |
Caissa 1.20 JA | 23.5/49 | -2 | 49 Partii |
Seer 2.8.0 | 30/48 | +12 | 48 Partii |
Viridithas 13.0.0 | 28.5/46 | +11 | 46 Partii |
Stockfish 16.1 | 18/44 | -8 | 44 Partii |
Clover 6.2 | 23/40 | +6 | 40 Partii |
JigSaw 5.9 | 17/38 | -4 | 38 Partii |
Motor 0.7.0 | 19.5/37 | +2 | 37 Partii |
Clover 7.0 JA | 22.5/36 | +9 | 36 Partii |
Peacekeeper 3.01 | 21.5/35 | +8 | 35 Partii |
Caissa 1.20 | 17/32 | +2 | 32 Partii |
Lizard 10.4 | 19/31 | +7 | 31 Partii |
Viridithas 14.0.0 | 17/29 | +5 | 29 Partii |
Yuliana 5.0 | 11.5/28 | -5 | 28 Partii |
Stormphrax | 16.5/26 | +7 | 26 Partii |
Stockfish 20240803 | 8/26 | -10 | 26 Partii |
Koivisto 9.2 | 16.5/25 | +8 | 25 Partii |
Clover 7.1 JA | 14/25 | +3 | 25 Partii |
Motor 0.6.0 | 13.5/19 | +8 | 19 Partii |
Neural network
Obsidian evaluates positions with a neural network trained on Lc0 data.
Comments
Post a Comment