Sturddle 2 is a fork of Sturddle Chess Engine (UCI) and neural network.
Author: Cristian (Cristivlas)
Building the Engine
The executable bundles binary images that support AVX512, AVX2, and generic SSE2.
Clang is recommended, the GNU C++ compiler may work but it is not supported.
If built with the `NATIVE_UCI` flag, invoke `main.py` to run the UCI engine.
Without the `NATIVE_UCI` flag, run `sturddle.py` instead.
## Training the Neural Net
I have trained the neural net on a large dataset that I generated and curated over a couple of years.
Sturddle 2.5 what's new?
Expected improved over 2.4.0 (TC: 12min+8sec)
Elo difference: +27.15 +/- 12.28
95% confidence: [+14.87, +39.43]
LOS: 100.00%
Total games: 1013
Rank Player Score Games W D L Win% Draw%
1 sturddle-2.5.0 546.0 1013 167 758 88 16.5 74.8
2 sturddle-2.4.0 467.0 1013 88 758 167 8.7 74.8
Added AVX512 inference kernel optimized for Zen 4 and Cooper Lake processors.
Revised architecture and retrained neural net.
Revised attack tables (magic bits) to use PEXT where available.
Optimized retry logic.
Sturddle 2.4 vs other engines:
| Quirky 2.0 | 12.5/22 | +3 | 22 Games |
| Zangdar 4.04.01 | 6.5/22 | -9 | 22 Games |
| Reckless 0.8.0 | 2.5/22 | -17 | 22 Games |
| Maelstrom 3.3.0 | 10.5/21 | +0 | 21 Games |
| Halogen 15.0.0 | 2/20 | -16 | 20 Games |
| Claw 2.0-M | 2.5/16 | -11 | 16 Games |
| Revolution 1.0 2708225 | 1/16 | -14 | 16 Games |
| Benbot 1.4.4 | 14/14 | +14 | 14 Games |
| Critter 1.6a | 9/14 | +4 | 14 Games |
| Invictus r391 JA | 8.5/14 | +3 | 14 Games |
| Cataphract 1.1 | 4.5/10 | -1 | 10 Games |
| Pawnocchio 1.8.1 | 1/10 | -8 | 10 Games |
| Potential 3.0.0 | 7.5/9 | +6 | 9 Games |
| Laser 1.8 beta JA | 4/8 | +0 | 8 Games |
| DarkZoryn 2.0 | 2/8 | -4 | 8 Games |
| Stockfish dev-20250902 | 1.5/8 | -5 | 8 Games |
| CorChess 20250824 | 1.5/8 | -5 | 8 Games |
| Spirit 300825 | 1.5/8 | -5 | 8 Games |
| OmegaCor 270825 | 1.5/8 | -5 | 8 Games |
| SF SE-240825 | 1/8 | -6 | 8 Games |
| Caissa 1.23 | 1/8 | -6 | 8 Games |

Comments
Post a Comment