Sturddle 2 is a fork of Sturddle Chess Engine (UCI) and neural network.
Author: Cristian (Cristivlas)
Building the Engine
<code>python3 tools\build.py</code> builds a native executable for the host OS. 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 2.3 what's new?
Proper SEE implementation.
Revisited LMR formula.
Increased NN accumulator size, added residual connection.
Optimized inference for Zen3/4+AVX512, Alderlake/Raptorlake+VNNI, support old Sandybridge processors.
Disabled end table probing.
Fix stack size in Windows binaries.
Sturddle2 2.01 vs other engines:
| Myrddin 0.92 | 22/38 | +6 | 38 Games |
| Mayhem 8.5 | 35/36 | +34 | 36 Games |
| Zangdar 2.31.04 | 11.5/36 | -13 | 36 Games |
| Monty 20241209 | 7/34 | -20 | 34 Games |
| RukChess 4.0.0 | 9/32 | -14 | 32 Games |
| Obsidian 14.13 SE | 1/32 | -30 | 32 Games |
| Lizard 11.1.5 JA | 8/31 | -15 | 31 Games |
| Critter 1.6a | 16/30 | +2 | 30 Games |
| Munchess 0.1 | 6/6 | +6 | 6 Games |
| Polonium 0.1.5 JA | 6/6 | +6 | 6 Games |
| Jackal 1.2.0 | 4.5/6 | +3 | 6 Games |
| Maestro 1.2.1 JA | 1.5/5 | -2 | 5 Games |
| Dog 2.5 JA | 4/4 | +4 | 4 Games |
| Clozza 1.0 | 4/4 | +4 | 4 Games |
| Nectar 0.2.1 | 4/4 | +4 | 4 Games |
| Rudim 1.4 JA | 4/4 | +4 | 4 Games |
| Lynx 1.8.0 | 2.5/4 | +1 | 4 Games |
| Illumina 2.0 | 2.5/4 | +1 | 4 Games |
| RukChess 4.0.1 | 1.5/4 | -1 | 4 Games |
| Mcu-max 1.0.5 | 3/3 | +3 | 3 Games |
| Crab 3.12.0 | 3/3 | +3 | 3 Games |
Comments
Post a Comment