Lc0 30xxx continues below the 10xxx series

Discussion of computer chess matches and engine tournaments.

Moderators: hgm, Harvey Williamson, bob

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Post Reply
Javier Ros
Posts: 109
Joined: Fri Oct 12, 2012 10:48 am
Location: Seville (SPAIN)
Full name: Javier Ros

Lc0 30xxx continues below the 10xxx series

Post by Javier Ros » Sat Dec 08, 2018 3:34 pm

The series 30 of Leela Chess Zero continues advancing but it does not reach the level of the series 10.
I have played two matches to compare:

Lc0_11261-Stockfish9
Lc0_31696-Stockfish9

Lc0 has used a GTX1060 GPU with 3Gb while Stockfish 9 has used three cores of the i7 4790S resulting in an approximate LeR ratio of 0.75. Lc0 played with endgame tabletables of 6 pieces, Stockfish9 no.
The time control was 25 minutes for the entire game with an increase of 10 seconds.
I used the Silver Suite of openings, that I think is the best one for testing with 50 well-chosen positions.

While version 11261 of Lc0 won Stockfish9 by a very small difference of 51.5%, Stockfish9 won convincingly Lc0_31696 with 27 wins and 59.5%

Lc019rc4_11261TB 100 (+ 13, = 77, - 10), 51.5%
Stockfish_9_x3 100 (+ 10, = 77, - 13), 48.5%

Stockfish_9_x3 100 (+ 27, = 65, - 8), 59.5%
Lc019_31696TB 100 (+ 8, = 65, - 27), 40.5%

The games of the series 30 were a hard fight where the aggressiveness of Lc0_31696 turned against it many times.

Here are all the games in pgn
https://sites.google.com/site/unoallavo ... ects=0&d=1

https://sites.google.com/site/unoallavo ... ects=0&d=1

The test of the series 10 was a complete disaster because release candidate lc0.19 hung several times, time forfeits, disconnected engines, etc, that forced to repeat many games.

I will comment some games of the series 30. In this testing there were no problems with final version 0.19.

In the first, a failed exchange sacrifice of Lc031696 gave victory to Sf9.



In the following game Lc031696 had a materialistic behavior not seen in series 10 and Stockfish9 finished the game dramatically.




Lc0 also attacked aggressively, as in the next with entry in seventh row and pawn coronation included.



In the following was Stockfish9 who got a pawn in sixth row and won



In the next Lc0 he won with one of his thorn pawns in sixth row.




This time it was Sf9 who arrived to sixth and seventh rows with pawns in the opposite castling, announcing mate in 10 on move 35.




Another wild attack of Sf9 with pawns on the castling of Lc0, with announcement of mate in 17 on move 41.



Sf9 announces mate in 15 on move 36 with another pawn in sixth! g6 +

The love relationship between a chess engine tester and his computer can be summarized in one sentence:
Until heat do us part.

User avatar
Guenther
Posts: 2511
Joined: Wed Oct 01, 2008 4:33 am
Location: Regensburg, Germany
Full name: Guenther Simon
Contact:

Re: Lc0 30xxx continues below the 10xxx series

Post by Guenther » Sat Dec 08, 2018 7:19 pm

Javier Ros wrote:
Sat Dec 08, 2018 3:34 pm
...

Lc0 played with endgame tabletables of 6 pieces, Stockfish9 no.

...

In the first, a failed exchange sacrifice of Lc031696 gave victory to Sf9.

Are you sure you played with all 6men for LC0?
Normally it should be impossible then to lose the 6men endgame which was on board since move 44 by Black?
Ok, giving away the pawn by 43.a7?! was already unnecessary, except you know the endgame is drawn anyway,
which should have been the case though.

Of course it can be that the implementation is buggy...

BTW your pgn is full of wrong three time repetitions. (same positions but with different right to move!)
Does this come from your used GUI or the pgn script used here in talkchess?

Javier Ros
Posts: 109
Joined: Fri Oct 12, 2012 10:48 am
Location: Seville (SPAIN)
Full name: Javier Ros

Re: Lc0 30xxx continues below the 10xxx series

Post by Javier Ros » Sat Dec 08, 2018 8:24 pm

Guenther wrote:
Sat Dec 08, 2018 7:19 pm


Lc0 played with endgame tabletables of 6 pieces, Stockfish9 no.


Are you sure you played with all 6men for LC0?
Normally it should be impossible then to lose the 6men endgame which was on board since move 44 by Black?
Ok, giving away the pawn by 43.a7?! was already unnecessary, except you know the endgame is drawn anyway,
which should have been the case though.

Of course it can be that the implementation is buggy...

BTW your pgn is full of wrong three time repetitions. (same positions but with different right to move!)
Does this come from your used GUI or the pgn script used here in talkchess?
Yes, the 6 men was in the directory and the Gui Arena was counting tablebases hits, but less than I expected.

The three time repetition were added by Arena and I don't how to avoid them.
The love relationship between a chess engine tester and his computer can be summarized in one sentence:
Until heat do us part.

Post Reply