Today I proudly present my new engine, crownd. There's always more to do and improve, but I have decided to finally publish it in this state, as I have reached I think a respectable strength.
Based on my own testing against different engines (on a ryzen 9 9950x), I estimate it around 3750 elo on the CCRL Blitz list. It seems to be strongest at short time controls, but it's still decent at longer time controls. I hope to get it tested and listed on CCRL, and that I will not get a bad surprise on its rating
The network is trained on Lc0 data, although my plan is to generate my own data from scratch. I have already experimented with it, but gave up on it for the first release because of the machine time needed that I was not able to give for now, as it was fully used for NNUE trainings, SPRTs and SPSAs, and I wanted to release it now.
For more info about its implementation, you can check out the README on github:
https://github.com/thomashuijbregts/Crownd
If you want to test it, I have published different versions for both Windows and Linux.
---------
I have come up with a (hopefully) new reduction idea, called "stability reductions" with positive results in my engine. The idea is to decrement the current depth if the static evaluation has been stable, within a margin, for a given amount of plies.
---------
AI disclosure: no AI written code - SIMD signatures provided by Claude