Trying out the Leela Hybrid engine (NN with AB)

[yanquis1972]

do you mean lower ST (if you're referring to the threshold value)?

[Laskos]

I think it's important to keep in mind that we assume Auxiliary engine must just help Lc0 tactically, and ideally it's much weaker than both Lc0 and its opponent. Many might be be tempted to use Auxiliary engine the level of SF, while their Lc0 is weaker than SF in some conditions. Even on my powerful GPU set-up this might occur. So, I took as Auxiliary engine Arasan 19.01 on 1 thread, an AB engine some 800-900 Elo points weaker than both Lc0 t30 and SF on 4 cores. It still helps a lot, solving many more positions on trimmed WAC tactical test suite and not harming almost at all the positional abilities of Lc0 (on my positional test-suite). But I changed some settings to get such a result with a weak Auxiliary engine:

AuxEngineBoost=20.00
AuxEngineDepth=10
AuxEngineThreshold=10
AuxEngineFollowPvDepth=3

In games, it increases Lc0 t30 strength against SF10 (4 cores) by some 35-40 Elo points, and the weak Arasan 19.01 on 1 thread helps almost as much as SF10 as an auxiliary engine (SF10 adds some 40-50 Elo points). The games were at bullet TC.

With SF10 as an auxiliary engine, many settings will work, as it is close in strength or stronger than Lc0, but with a weak auxiliary AB engine, finding correct settings is trickier. It has to know when it can intervene, and this is usually the case of gross tactical mistakes of Lc0. Still, endgames are bad.

[Eduard]

Why is it not possible for the AuxEngine to use endgame tablebases?

[Eduard]

I see that it is also possible to combine two neural networks. One as the main engine and one as the AuxEngine. Interesting.

[glennsamuel32]

I'm not a C++ programmer or an expert on chess engines, so I tried to keep it simple.

Thanks, super effort for a first attempt !!
Works fine without any issues for me on Fritz 15, even with ponder on for both engines.

Could you elaborate on what "AuxEngineFollowPvDepth" means ?

[WinPooh]

I see that it is also possible to combine two neural networks. One as the main engine and one as the AuxEngine. Interesting.

And then add a third NN to decide which one to use.
But... isn't it equivalent to having just one, larger, network?

[jjoshua2]

I see that it is also possible to combine two neural networks. One as the main engine and one as the AuxEngine. Interesting.

I tried that and it seemed to work for me.

[KillerDucky]

v0.21.0-fish-v0.4 released! Just one feature addition and one minor bug fix.

* Add support for AuxEngine to use Syzygy endgame tablebases. Automagic if you configure Lc0 to use Syzygy and AuxEngine supports them.
* Fix a bug with boosting Policy to larger than 100%. Also set maximum AuxEngineBoost to 100%. Internally Lc0 uses a compressed format for Policy so values larger than 100% do not work.

[sincerelybrent]

Terrific! Have a 0.3 gauntlet going atm. Will download soon. Thanks for all of your hard work on this, KillerD

[supersharp77]

LeelaFish is a modified version of the Lc0 engine that gets help from an outside (auxiliary) engine. Leela uses the AuxEngine's PV to boost the Policy of those moves.

https://github.com/killerducky/lc0/wiki/LeelaFish

Downloaded it from the github site yesterday tried to get it to work this morning...got a cuda dll error with one LC0 compile and my AVG (antivirus) ate the second compile and said "Will check exe with the home (AVG) site and get back to me"
haven't heard back from AVG all day...anyone get it to work in cpu? Thx AR