I'm Michael, from the UK, and I've been reading here for a while. I'm writing a chess engine called Cadence, in Rust: github.com/grounzero/cadence
Where it is today:
- Alpha-beta search with the usual modern pruning and reductions, and a deliberately simple hand-crafted evaluation for now
- Every change is tested by SPRT on my own OpenBench instance, and kept only if it passes
- It plays on Lichess as BOT Mithandros, running on a Jetson Orin Nano
One question for people who have made that switch: for a first network trained on data from a hand-crafted engine, how many positions did you find you really needed before the net beat your HCE comfortably?
Thanks, and I'm looking forward to learning from you all.
Michael