Schoenemann - UCI Chess Engine
Author: Jochengehtab
This is the first release of Schoenemann.
It uses a NNUE with 512 HiddenLayers and PVS as it's search function.
Testet against Akimbo v0.4.0 which has roughly shown that the engine is around 150 elo weaker (at LTC and STC).
It does currently not support FRC.
It uses a NNUE with 512 HiddenLayers and PVS as it's search function.
Testet against Akimbo v0.4.0 which has roughly shown that the engine is around 150 elo weaker (at LTC and STC).
It does currently not support FRC.
v.0.4.0:
A long time has passed since the last release. In this version of Schoenemann, I wanted to switch from the Leela Net to a selfgen Net. But as it turns out, my self-gen nets are far behind (around 170-200 Elo) the Leela Net. But nevertheless, to try something new, I used one of the first persons, Montydata (https://github.com/official-monty/Monty), as the training data, which is not so far behind the Leela Net compared to my selfgen Net. Also, I made some improvements in the search:
Speedup (with march=native)
Quiet History
Continuation History
Cutnode with LMR
Improving over multiple plies
LTC Tune
I will continue to train the selfgen Net, but this takes much more resources. So hopefully I will get a good selfgen Net in the next few releases. And thanks to everyone who supported me, especially the Fury-Bench members! (Thank you)
And here are some progress tests:
(Against master)
Elo | 323.42 +- 24.32 (95%)
Conf | 8.0+0.08s Threads=1 Hash=16MB
Games | N: 1000 W: 771 L: 40 D: 189
Penta | [0, 6, 44, 163, 287]
https://chess.aronpetkovski.com/test/6897/
Elo | 283.58 +- 18.99 (95%)
Conf | 40.0+0.40s Threads=1 Hash=16MB
Games | N: 1000 W: 705 L: 32 D: 263
Penta | [0, 1, 48, 228, 223]
https://chess.aronpetkovski.com/test/6898/
(Against Integral v4)
Elo | -16.85 +- 15.49 (95%)
Conf | 8.0+0.08s Threads=1 Hash=16MB
Games | N: 1032 W: 325 L: 375 D: 332
Penta | [47, 123, 214, 97, 35]
Comments
Post a Comment