Willow - UCI engines Rating CEDR=3381
A chess engine I am currently working on, written in C (WORK IN PROGRESS)
Compiled in Windows, works on BanksiaGUI with limited UCI support right now.
A chess engine I am currently working on, written in C (WORK IN PROGRESS)
Compiled in Windows, works on BanksiaGUI with limited UCI support right now.
v.4.0:
Willow's final version for now, 4.0 is a massive leap in strength from 3.1, with an estimated CCRL elo of 3500+ in blitz and 3400+ in 40/15.
Score of 4.0 vs 3.1:
Elo | 176.41 +- 15.06 (95%)
Conf | 60.0+0.60s Threads=1 Hash=128MB
Games | N: 1006 W: 491 L: 20 D: 495
Penta | [0, 10, 121, 263, 109]
https://antares2262.pythonanywhere.com/test/71/
Binaries:
v2 (for old computers) and v3 (should work for most computers) versions of Willow for both Windows and Linux are included. If you have a computer that supports AVX512, or none of these binaries work on your machine, you can build from source with these commands:
git clone https://github.com/Adam-Kulju/Willow
cd Willow
make
and the binary "willow" (or "willow.exe" on Windows) will be in the same directory.
New features:
a new bucketed net (768*4 -> 768 -> 1) trained on 2b self-play positions
many search improvements (such as negative extensions, nmp eval conditions, minor things like that)
an SPSA run of my search function on 5k iterations
SIMD
completely fixed multithreading
An announcement:
Willow now plays at TCEC and will soon play at CCC, and I am so proud of it. It's the greatest project of my life so far. However, further improvement is becoming difficult and more importantly I have strayed away from my original goal of creating an exceptionally aggressive engine that is also superhuman. In the chase for Elo, I abandoned the "intentional mistakes" part of datagen because it made the engine objectively stronger at the expense of style. Indeed, I rarely watch its games anymore and while it still seems to be reasonably aggressive and can certainly mow down weaker opponents in style, I can't in good faith call it an aggressive engine anymore. For this reason I am moving on to another project that I will release and advertise when it is ready. Until then, it has been a pleasure being part of the amazing community of engine devs out there!
Acknowledgements:
As always a huge thank you to my SweHosting friends. I would have stalled at 2500 if not for you. You guys are the best.
Willow 3.1 results:
| Texel 1.10 | 9.5/22 | -3 | 22 Partii |
| Akimbo 0.6.0 | 11.5/20 | +3 | 20 Partii |
| Stormphrax 2.0.0 | 9/20 | -2 | 20 Partii |
| Carp 3.0.1 | 5.5/16 | -5 | 16 Partii |
| Alexandria 5.0.0 | 4/16 | -8 | 16 Partii |
| Luna 2.0.0 | 14/14 | +14 | 14 Partii |
| Princhess 0.13.0 | 13.5/14 | +13 | 14 Partii |
| Altair 5.0.0 | 11/14 | +8 | 14 Partii |
| Viridithas 11.0.0 | 2.5/14 | -9 | 14 Partii |
| Mida 2.1 | 11.5/12 | +11 | 12 Partii |
| Demolito 2023-08-20 | 9.5/12 | +7 | 12 Partii |
| Avalanche 2.0.0 | 6/12 | +0 | 12 Partii |
| Yuliirma 2.1 | 0.5/12 | -11 | 12 Partii |
| AbbyStein 1.5 | 0.5/12 | -11 | 12 Partii |
| Hazard 2.06 | 0/12 | -12 | 12 Partii |
| Cool Iris 11.30 | 0/12 | -12 | 12 Partii |
| Obsidian 4.0 | 8/10 | +6 | 10 Partii |
| Midnight 8 | 7/10 | +4 | 10 Partii |
| Tofiks 1.2.0 | 8/8 | +8 | 8 Partii |
| Caissa 1.13.1 | 2/8 | -4 | 8 Partii |
| RofChade 3.1 | 1.5/8 | -5 | 8 Partii |
Comments
Post a Comment