Stefan Pohl Computer Chess

private website for chessengine-tests


NN longtime testing (SALC Armageddon Chess)

 

SALC Armageddon means: White is allowed to castle short, only and Black is allowed to castle long, only (=SALC (Short And Long Castling)). And all draws are counted as a win for Black (=Armageddon), because these castling-rights are a advantage for White. So, there are no draws anymore. And no draw-death of computerchess. Read more and download more Armageddon stuff in the "Armageddon openings"- section of my website.

 

Playing conditions:

 

Hardware: i7-8750H (Hexacore) Notebook with RTX 2060 GPU, Windows 10 64bit, 16GB RAM

CPU-Speed: Stockfish with 97% CPU-Speed (to switch off the Intel Turbo Boost): 7.5 MN/s in starting-position, running on 11 threads.

GPU (used by LC Zero): Nvidia RTX 2060 (6GB). LC Zero calculates around 11500 n/s in the starting position (I used the MSI-Afterburner-tool to reduce the speed of the RTX-Card as far as possible) (measured with "go infinte") with Net 32930 (Netsize 20x256), which means a Leela-Ratio (what is Leela Ratio? look here) of 1.3.The Leela-Ratio-value of AlphaZero (used a 20x256 net, too) in the match vs. Stockfish 8 was 1.0 - so 1.3 is a high value, but acceptable.

Hash: 4096 MByte for Stockfish and 5.000.000 size of NNCache for Leela

GUICutechess-cli (GUI ends game, when a 5-piece endgame is on the board)

TablebasesNone for engines, 5 Syzygy for cutechess-cli

Openings: 150 selected (and SALCed) Noomen 4 moves openings, download them here

Large Memory Pages: Off

Ponder: Off

Thinking time: 8' + 5'' (average game-duration: 30 minutes, a 300 games testrun takes around 6 days...)

 

LC Zero Github (information, download Networks and LC Zero Engine): here

Read more about, how LC Zero works, in the LC0-Blog: here

 

Download all played games (non-armageddonized) here

 

Latest update (2020/01/21): Fat Fritz 1.0. Complete disappointment, even though the thinking-time was long. Next testrun: Repeat testrun of Lc0 0.23.1 LS 12.2 with Kiudee-setting, which give measureable Elo-gain in shorter timecontrols in various testruns. Lets see, if there is an Elo-gain with long thinkingtime, too...

 

300 SALC-Armageddon games each testrun (=all draws are a win for Black)
vs Stockfish 191210 (SPCC-Elo: 3555)

1 Lc0 0.23.1 LS 12.2 (20x256)    : 3562 300 (+153,=  0,-147), 51.0 %
2 Lc0 0.23.1 t40-1705 (30x384)   : 3548 300 (+147,=  0,-153), 49.0 %
3 Lc0 0.23.1 42741 (20x256)      : 3541 300 (+144,=  0,-156), 48.0 %
4 Lc0 0.23.1 61894 (24x320)      : 3529 300 (+139,=  0,-161), 46.3 %
4 Lc0 0.23.1 J13B.2-200 (24x320) : 3529 300 (+139,=  0,-161), 46.3 %
5 Fat Fritz 1.0 (20x256)         : 3477 300 (+117,=  0,-183), 39.0 %


Games        : 1800 (finished)

White Wins   : 905 (50.3 %)
Black Wins   : 895 (49.7 %)
Draws        : 0 (0.0 %)
Unfinished   : 0

White Score  : 50.3 %
Black Score  : 49.7 %

 

"k"-extension in Lc0 engine-name means, Lc0 was running with Kiudee-Setting:

CPuct = 2.147

Fpu = 0.443

Policy Temperature = 1.607

CPuctBase = 18368

CPuctFactor= 2.815