I thought I was done with chess programming, but I made the mistake of glancing at TalkChess a few months ago and the posts of a person with seemingly boundless energy and enthusiasm called Maksim Korz kickstarted me again; damn you Maksim - now I'm addicted again
Lozza 2.0 is pretty much unchanged except that material, PSTs and all evaluation params were tuned in one go using Peter O's 'Texel' method. The tuning code is in the repo if any other Javascript developers are interested. It's 'the' +1/-1 slow algorithm, but I was not in a rush and It ran for weeks!
op12no2 wrote: ↑Thu Feb 18, 2021 11:28 am
Cheers Tamás; tuning was pretty interesting. But if I had known about NNUE when I started maybe I would have skipped it
I haven't dropped the old tuned evaluation yet.
Maybe I won't. Even so, my Neural Network works.
Unfortunately, I have little time to experiment further now.
tomitank wrote: ↑Thu Feb 18, 2021 12:42 pm
I haven't dropped the old tuned evaluation yet.
Maybe I won't. Even so, my Neural Network works.
Unfortunately, I have little time to experiment further now.
I was surprised what tuning gave in self play; but whether or not it translates to gains against other engines and longer time controls; only time will tell.
Yeah, while I was catching up I noticed that you have other commitments now - and also that your progress on Tomitank over the last few years has been very impressive. So much so that I didn't even consider testing against it