Trying out the Leela Hybrid engine (NN with AB)

[corres]

I have made a binary from a modified source of Leelafish in which I used fixed move time instead of fixed depth.
My experience is using too much Stockfish it does not enhance the Elo of Leelafish but it decreases it.
I think there is a great difference in style between Stockfish and Leela and because of this too much Stockfish disturbs the own evaluation of Leela. To find an optimum where the gain of Elo is on its maximum is not an easy task because it depends on a lot of parameters, mainly from the thinking time, duration of the game and the parameters of the helper engine.

[Gregory Owett]

For you, what is the best auxiliary engine ?

[sincerelybrent]

For you, what is the best auxiliary engine ?

So far, Houdini 6 w/ "tactical mode=1" > Stockfishes. Haven't tried CFish, yet, but I'm going to soon.

[Gregory Owett]

Thanks, I will try them!

[KillerDucky]

v0.21.0-fish-v0.6 released!

Note the new name AuxEngineFile, you'll need to update your GUI to use that option instead of the old one.

* Rename AuxEnginePath to AuxEngineFile (helps chess GUIs know to select a file not a directory).
* Fix some crashes parsing AuxEngine pv output.

I tested a few engines and verified they're fixed with this release. Please let me know any engines that still don't work (including a log file with AuxEngineVerbosity=10 would be helpful too).

[Eduard]

With CPU-version Komodo works now, thank you!

[CositasBuenas]

Windows Defender Antivirus raises the alarm and refuses the download (v0.6). What to do?
Michael

[CositasBuenas]

I think there is a memory leak. Would you please check this!
Regards, Michael

[dkappe]

Some thoughts:

- Just when you want to see if leelafish prevents a “brilliant positional sacrifice,” you can’t lay your hands on one. Has anyone got a collection of leela positional stunners they can test against?
- lowering threshold means many more positions will be looked at. Since the aux engine is the bottleneck, you need to lower the depth to process them all, or have several engine instances running in parallel.
- there’s a paper that sought to guide mcts with ab search (https://dke.maastrichtuniversity.nl/m.w ... _final.pdf). As I recall, 3 ply was the winner. Now there was no nn guiding the mcts, but I’m curious if depth 3 and threshold 1 (or 0) does better than 15/100 or if the truth lies somewhere in between.

[corres]

Some thoughts:

- Just when you want to see if leelafish prevents a “brilliant positional sacrifice,” you can’t lay your hands on one. Has anyone got a collection of leela positional stunners they can test against?
- lowering threshold means many more positions will be looked at. Since the aux engine is the bottleneck, you need to lower the depth to process them all, or have several engine instances running in parallel.
- there’s a paper that sought to guide mcts with ab search (https://dke.maastrichtuniversity.nl/m.w ... _final.pdf). As I recall, 3 ply was the winner. Now there was no nn guiding the mcts, but I’m curious if depth 3 and threshold 1 (or 0) does better than 15/100 or if the truth lies somewhere in between.

In this case Stockfish is not the optimal choice for Leelafish because Stockfish works as a Stockfish
if depth > 12.
Maybe we need a specialized Stockfish for Leelafish, I think.
But this question can be decided by tests only.