Re: Three fishes vs Leela
Posted: Thu May 23, 2019 9:44 pm
It might depend on the hardware used. If I use on one hand Leela on RTX 2070 GPU + 2 cores and on the other SF on 2 cores (I have 4 fast cores CPU), Leela is surely to be preferred as the main engine. Not only it is very much stronger, but aside tactical blunders in complicated positions and some endgames, Leela's moves are almost universally better than SF ones. It seems you want to analyze already played games forward and backward using the regular hash scheme of SF, and you even have some "time to solution", are you checking for blunders? In that case it's better to use SF without any Leela. Strength-wise and say correspondence and centaur chess, in my conditions, Leela is clearly the main engine having the say in close to 90% of the moves. LeelaFish still seems to not be able to correct some endgames (just an impression) during the gameplay (with Leela being the main engine).peter wrote: ↑Thu May 23, 2019 8:09 pmSo, in case of a Fish as aux-engine it should be called Fishleela, shouldn't it?Uri Blass wrote: ↑Thu May 23, 2019 6:59 pm I do not understand the name leelafish.
https://github.com/killerducky/lc0/wiki/LeelaFish
LeelaFish is a modified version of the Lc0 engine that gets help from an outside (auxiliary) engine.
There is Lc0 and another engine and the second engine does not have to be stockfish so why the name leelafish?
My trials with combinations of Fish and Leela show much better effect at supporting SF with Leela instead of other way round, because SF is much better at hash-learning, so if moves differ, SF is to be brought to change its mind much better as for getting points of moves and lines given in from outside of own pondering. One could argue, that might be because of moves from Leela being better more often, but that doesn't seem to be so to me as for evaluating moves along longer lines at Forward- Backward. To say it shortly, backward analysis with SF is much more effective than with Leela, if you compare time to solution and time to eval.
To talk about it relevantly, we would have to discuss certain positions always only of course.
You see that best at interactive analysis with using both engines, which of course isn't to be measured Elo- wise except at corr.-chess, but as for Nucleus I do prefer SF as master engine too. (LC0 wouldn't have to give first ply- suggestions at all anyhow as master engine in Nucleus- combi.)