TalkChess.com

Getting a smooth switch is still a bit tricky. Still, the results are promising. Here a game between FF+LE and 42850 at 3 seconds per move on a 1070. Using Noomen 3. Switch is at 16 men. [Event "?"] [Site "?"] [Date "????.??.??"] [Round "3"] [White "FF+LE"] [Black "42850"] [Result "1-0"] 1. d4 Nf6 2...

Use ful for naive comparison of testing configurations. Use less as a basis for determining what is balanced. Bingo! When I came up with the “leela ratio” (a name that I now regret), the whole focus was around beating A0. As t10 started getting good, people started posting breathless match results ...

Scorpio just lost in a real howler to Komodo at CCCC. Here the key position: 8/2n5/6p1/P2Pk3/6Kp/7P/6P1/8 w - - 0 1 Here Scorpio played Kxh4?? I was able to reproduce this, but the amazing thing is that almost all of the leela nets I tried also played this howler —- 11258, 42850, t60, LD2, t58, Bad ...

According to the testers on discord, the T58 run has passed LD2 in strength. My own endgame tests put 58432 in third place behind LE and 42850. Looks like there’s more playing strength to be squeezed out of even the smaller nets.

Are we even talking about the same Sergio?

Some people say it’s 10-20 elo weaker than 42850, some say it’s about even. I can find tests in the leela chess discord on both sides of this issue. In most cases, various late t40 nets and t40 derived nets are within a few points of each other.

Marek Soszynski wrote: ↑

Sun Nov 24, 2019 10:26 pm

dkappe wrote: ↑

Sun Nov 24, 2019 10:17 pm

I’m pretty sure a Fat Fritz jr for CPU is on the way.

Slim Fritz?

That’s a German blues guitarist from Hannover.

You can also use it (Fat Fritz) in another GUI if Chessbase is not your thing. No, I cannot :) At least not on a GPU because I don't have CUDA and OpenCL doesn't work, and I already have weights optimized for CPU, so I didn't bother. In their defense, I'm the only person around that I know that has...

corres wrote: ↑

Sun Nov 24, 2019 8:17 pm

dkappe wrote: ↑

Sun Nov 24, 2019 7:33 pm

I think various cpuct settings and policy softmax temp are also different.

I EVERY parameters of Fat Fritz modified as they are in Leela.

I don’t understand. You did or did not change the cpuctbase, cpuctfactor, policy softmax temp to match those in the Fritz 17 executable?

I think various cpuct settings and policy softmax temp are also different.