It has been established a long time ago that Stockfish is better than Lc0 in almost all categories. Leela needs more training but by then, Stockfish has gained even more ELO. A more fair match would be to put Stockfish on a weaker CPU such as a Ryzen 1600 and Leela on a 2080ti.
@Vetinari_Computer The games from the AlphaZero paper you're referring to I believe had asymmetric time controls, in the sense that Stockfish had more time than AlphaZero, and they were testing how AlphaZero performed at a time disadvantage.
In the TCEC, this season, there was blitz testing in Div3 [where LCZero was before it promoted], and its performance in the 5+2 games for the blitz testing was brutally stronger than its performance in the 30+10 games for the division.
In the TCEC, this season, there was blitz testing in Div3 [where LCZero was before it promoted], and its performance in the 5+2 games for the blitz testing was brutally stronger than its performance in the 30+10 games for the division.
@JohnJPershing no it wasn't time odds or anything. Check out arxiv.org/pdf/1712.01815.pdf page 7. As you can see Alpha 0 scales a lot better than Stockfish. (as well as better than Elmo in Shogi)
EDIT: Well it was time odds, but it was symmetrical, they tested that with both engines not just one.
EDIT: Well it was time odds, but it was symmetrical, they tested that with both engines not just one.
@Vetinari_Computer Hmmm. There was another section on asymmetric time odds, but this appears damning. Perhaps there needs to be some analysis of LCZero vs AlphaZero and look at their differences. I'll definitely have to read up more on the subject.
Thanks for the source, I appreciate it.
Thanks for the source, I appreciate it.
Just curious why only 6 games. It is understandable why humans can't play 30 pro games per day, but it is not hard for two engines to play the whole day. Lichess just need to implement an api for an engine to challenge another engine.
Niko, the owner of stockfish or Leela probably need to use those machine in something else. Plz enjoy whatever freebies you get.
6 games lol.
And yes, stockfish scales better than a0. If you watch not that bogus figures from paper but carefully read what a0 paper says (the newest one) you will see that a0 gains like 10 elo per doubling time at VVLTC, stockfish gains 20-30.
And all this "better scaling" of a0 was claimed because it's ultrabad in blitz.
And yes, stockfish scales better than a0. If you watch not that bogus figures from paper but carefully read what a0 paper says (the newest one) you will see that a0 gains like 10 elo per doubling time at VVLTC, stockfish gains 20-30.
And all this "better scaling" of a0 was claimed because it's ultrabad in blitz.
Also, why is it sf10? It should be sf 11 dev if you want to see the highest level of chess
fyi, sf dev builds are on http://abrok.eu/stockfish/
Why will play Lc0 v19.1, and not a newer version with significant changes?
This topic has been archived and can no longer be replied to.