Leelafish with distilled network

[Eduard]

Here are 3 interesting analyzes. Leelafish used the same main Engine (network 11258 64x6) and AuxEngine (Cfish 230219).
I just wondered why Leelafish gets deeper than Leela?

Lc0 only CPU.
GUI=Fritz 15.

[OPTIONS]
Threads=3
AuxEnginePath=C:\Program Files (x86)\ChessBase\Engines\lc0-v0.21.0-fish-v0.5-distilled\Cfish.exe
AuxEngineOptions=Hash=128;Threads=1
AuxEngineThreshold=45
AuxEngineBoost=86.000000
AuxEngineDepth=15
AuxEngineFollowPvDepth=5

Caruana,F - Carlsen,M 1/2-1/2, World Chess Championship 2018
[d]1r1q1rk1/1p2b2p/pN6/P1pP1bp1/5p2/2B2Q2/1P4PP/R4RK1 w - - 0 1

After 5 minutes:

Analysis by Lc0 v0.21.0-rc1se (11258 64x6):

23.Tad1 Ld6 24.Sc4 g4 25.Df2 f3 26.Tfe1 Tc8 27.Sxd6 Dxd6 28.gxf3 Tce8 29.Txe8 Txe8 30.fxg4 Lxg4 31.Td3 Te2 32.Dg3 Dg6 33.h3 h5 34.hxg4 Dxg4 35.d6 Tg2+ 36.Kxg2
+/- (1.33) Tiefe: 12/29 00:05:02 148kN
23.Tad1 Ld6 24.Sc4 g4 25.Df2 f3 26.Tfe1 Tc8 27.Sxd6 Dxd6 28.gxf3 Tce8 29.Txe8 Txe8 30.fxg4 Lxg4 31.Td3 Te2 32.Dg3 Dg6 33.h3 h5 34.hxg4 Dxg4 35.d6 Tg2+ 36.Kxg2
+/- (1.33) Tiefe: 12/30 00:05:05 149kN

Analysis by Cfish 230219 64 POP N:

23.Tae1 Lf6 24.h4 Ld4+ 25.Lxd4 cxd4 26.hxg5 Dxg5 27.d6 Lg4 28.Db3+ Kh8 29.Db4 f3 30.Dxd4+ Df6 31.Te4 Dxd4+ 32.Txd4 Tg8 33.d7 fxg2 34.Tff4 h5 35.Sc4 Tbd8 36.Se5 Kh7 37.Td6 Lh3 38.Th4 Lg4 39.Tb6 Kg7 40.Sxg4 hxg4 41.Txb7 Kf8 42.Th6 Tg7 43.Th8+ Tg8 44.Txg8+ Kxg8 45.Tc7 Kf7
+/- (1.51 --) Tiefe: 36/69 00:04:47 930MN, tb=23028
23.Tae1 Lf6 24.h4 Ld4+ 25.Lxd4 cxd4 26.hxg5 Dxg5 27.d6 Lg4 28.Db3+ Kh8 29.Db4 f3 30.Dxd4+ Df6 31.Te4 Dxd4+ 32.Txd4 Tg8 33.d7 fxg2 34.Tff4 h5 35.Sc4 Tbd8 36.Se5 Kh7 37.Td6 Lh3 38.Th4 Lg4 39.Tb6 Kg7 40.Sxg4 hxg4 41.Txb7 Kf8 42.Th6 Tg7 43.Th8+ Tg8 44.Txg8+ Kxg8 45.Tc7 Kf7
+- (1.86 ++) Tiefe: 36/69 00:05:21 1044MN, tb=25308

Analysis by Leelafish v0.21.0-fish-v0.5se (11258 64x6 + Cfish 230219):

23.g4 Lg6 24.Tad1 Ld6 25.Tfe1 Dc7 26.Sc4 Tfe8 27.Te6 Txe6 28.dxe6 Td8 29.Sb6 c4 30.Sd5 Dc6 31.h4 Ld3 32.Kh2 Te8 33.Te1 Le7 34.hxg5 Lxg5 35.Lf6 Txe6 36.Txe6 Dxe6 37.Lxg5 Le2 38.Sf6+ Kf7 39.Dxf4 Ld3 40.Dc7+ Kg6
+- (1.86) Tiefe: 15/38 00:04:58 144kN
23.g4 Lg6 24.Tad1 Ld6 25.Tfe1 Dc7 26.Sc4 Tfe8 27.Te6 Tbd8 28.Lf6 Lf7 29.Sxd6 Txd6 30.Txe8+ Lxe8 31.Le5 Dd7 32.Lxd6 Dxd6 33.De4 Lf7 34.h4 gxh4 35.Kg2 h5 36.gxh5 Lxh5 37.Tg1 Kf8 38.Kh3 Lf7 39.Df5 Dxd5 40.Dxd5
+- (1.86) Tiefe: 15/38 00:05:03 146kN

Very last Analysis!

[Eduard]

Btw: With Cfish version "CF EXT 170219 x64 POP" (renamed as CF.exe) is it possible to enable a "tactical mode",
and disable "LazyThreshold", if I want. Thanks for this CF EXT.. possibilities.

AuxEngineOptions=Hash=128;Threads=1;Tactical=true;LazyThreshold=false;

Here for example a analysis:
[d]k1r4r/4q1p1/PBp2b1n/1pPppb1B/4p3/R6P/1PPQ1PP1/4NRK1 b - - 0 1

Analysis by Leelafish v0.21.0-fish-v0.5 (CF tactical):

26...Lg5 27.Dd1 Lf4 28.c4 bxc4 29.Sc2 d4 30.Sb4 d3 31.Lf3 exf3 32.Dxf3 Sg4 33.Sxc6 Dh4 34.g3 Dxh3 35.Se7+ Le4 36.Dxe4+
-/+ (-1.09) Tiefe: 10/25 00:01:55 48kN
26...Lg5 27.Dd1 Lf4 28.c4 bxc4 29.Sc2 d4 30.Te1 g6 31.Le2 d3 32.Sb4 dxe2 33.Dxe2 Sg4 34.Tea1 e3 35.fxe3 Lh2+ 36.Kh1
-/+ (-1.10) Tiefe: 11/25 00:01:58 49kN

[Damir]

Hi Eduard.

Where can I find this Leelafish you have been using ?

[dkappe]

I’m not completely sold on the playing style of leelafish, but you can’t argue with the strength. With Sting 14, depth 10, threshold 35, pv depth 3 and 11258-120x9-se, 1 thread all around, leelafish runs about even with the latest ethereal, a jump of 150 elo over vanilla lc0. This is with tc 6+6.

Code: Select all

Score of ethereal vs fish-120x9: 1 - 2 - 13  [0.469] 16

Looking at the cpu usage of Sting 14, it averages around 30% at these settings, so isn’t doing nearly as much as it could.

Jury is still out on the playing style.

[gerhard]

Hi Eduard.

Where can I find this Leelafish you have been using ?

https://github.com/killerducky/lc0/releases

[Eduard]

Sorry Damir, I thought you already knew it! No?

[Damir]

I did knew about it, but there was a user on Infinity who was getting 50.000 kN/s with his 36 cores with leelafish, so I assumed he was using a different kind of leelafish… cause the only Leelafish I know its kN/s are very low.

[Eduard]

you looked very carefully!

[Eduard]

At the moment I am testing the network 50405 with Leelafish on CPU only. It looks like this network is better than destilled networks! Previously I tested NN 50326, it was not bad but 50405 is much better! It seems to be the best NN for CPU only, it plays strong.

[Guenther]

At the moment I am testing the network 50405 with Leelafish on CPU only. It looks like this network is better than destilled networks! Previously I tested NN 50326, it was not bad but 50405 is much better! It seems to be the best NN for CPU only, it plays strong.

May be this is true for Leelafish, but with LC0 alone without any A/B helper program it has has a long way to go vs. very tactical engines.
It seems the tactical holes are filled only after an enormous lot of training games. Try for example Hannibal 1.7.

I tested already a few NN50s and noticed that they were partially en par with best dist. 11258 until they had to play against very tactical opponents.