Not logged inRybka Chess Community Forum
Up Topic The Rybka Lounge / Computer Chess / TCEC 14 superfinal opening PGN available
- - By Venator (Silver) Date 2019-02-24 12:16
The TCEC 14 superfinal opening PGN is available for download. As well as all the 100 games played. See:

This superfinal was an epic fight between reigning champion Stockfish and AlphaZero-based Leela Chess Zero. In the end Stockfish won by the narrowest possible margin: 50.5-49.5. Lots of brilliant games, blunders, drama, missed opportunities and the most exciting score I have ever seen: Stockfish took the lead 2-0, then it became 2-2, 3-2 and 3-3. Suddenly Stockfish won three in a row, making it 6-3. But Leela fought back, equalized and then ran away: 8-6 and then 9-7. But in the end Stockfish managed to turn things around: 10-9, with 81 draws.

Congrats to both team Stockfish and team Leela for a thrilling superfinal! It cost me a lot of sleep, but it was worth it :-).
Parent - - By Labyrinth (*****) Date 2019-02-24 23:52
Great to hear it! Excellent for computer chess.
Parent - - By Venator (Silver) Date 2019-02-27 08:26 Edited 2019-02-27 08:39
Agreed! The clash between two engines with a completely different style is very interesting. My analysis of the strengths and weaknesses, based on what I have seen sofar:

Leela Chess Zero: plays positionally better than SF; especially strong in openings and the middle game; has a lot better understanding of pawn structures compared to SF (and other AB engines); spots nice positional ideas that are missed by SF.

Weaknesses: misses tactical shots; endgame play is worse than SF; struggles to win won games; fails to win won games; often takes a far longer time to convert a winning endgame; often way too optimistic eval, even in drawn positions (f.e. where there is perpetual check).

Stockfish: the king of tactics; better in endgames than Leela; finishes off won games easily and quickly; better understanding of drawn positions, like f.e. perpetual check.

Weaknesses: overlooks nice positional ideas by Leela; often takes time to see through the positional problems that Leela poses; and by far the weakest point in SF (and AB engines): pawn structures. You'll see a lot of weak structures in SF's games, like weak pawns, blocked pawns, putting pawns on the same colour of its bishop, weakened colour complexes, the absence of pawn levers etc. SF fails to understand the importance of keeping a fluid pawn structure, either to protect squares, or to break open the position.

Strength of both: understanding king's attacks; brilliant defenders, even in precarious positions.
Parent - - By Venator (Silver) Date 2019-02-27 08:52
Another strange SF habit is that it doesn't care about 0.00 positions. To explain this a bit further: you have different types of drawn positions:

1. Strong draws, i.e. one side is pushing, the other side can only defend and sometimes have to make only moves to draw.
2. Weak draws, i.e. one side only can defend, has zero chances of winning and has the possibility to lose.
3. Even draws, i.e. it's a clear draw and no side can win.

I understand that for SF 'a draw is a draw'. From a human perspective this is a strange choice: of course from a winning perspective, you should always go for the strong draw and avoid weak draws. But SF often does the opposite: it goes from a strong draw into a weak draw. This makes no sense. You will never win a weak draw. At least in a strong draw there is a chance.

In the checkers program Chinook there was code to distinguish strong draws from weak draws. Leela does something similar, i.e. giving a positive score if it believes it has a small chance of winning and a negative score if it has a small chance of losing.
Parent - - By Dragon Mist (****) Date 2019-03-01 22:01

> But SF often does the opposite: it goes from a strong draw into a weak draw.

Yes, noticed that, too. If the position is drawn, it will, for instance, give away 3 pawns, still holding a draw - while before that it was the opposite side that was fighting for draw (some OCB endings, perpetual checks, and even some non elementary draws that it still recognizes). I tried to talk to people with developing skills involved in SF, but don't think anyone got this problem seriousy.
Parent - - By Uly (Gold) Date 2019-03-02 09:48
Have you tried ShashChess? It's a Stockfish flavor that seems to fix this issue. Not only that, but for positions that SF shows some 0.32 score for the top two moves, Shash (Petrosian) would give some 1.60 to one, 0.80 to the other, where 1.60 is the least drawish. Clear move preference while SF likes most moves equally.
Parent - By Dragon Mist (****) Date 2019-03-03 12:54

> Have you tried ShashChess?

No, I do not use various SF forks out of principle. Nice to hear there may be a solution to this.
Parent - - By gsgs (***) Date 2019-02-25 06:43 Edited 2019-02-25 06:47
StockFish with White gave +0.60 in average to the openings
but only 0.10 with Black due to contempt

Lc0 gave +0.80 and +0.70

evals :
chart :
Parent - - By Venator (Silver) Date 2019-03-06 08:48
Thanks. After the TCEC 14 superfinal there was a bookless bonus match between SF and Leela, with TC = 12'+3". The result:

Leela Chess Zero - Stockfish 56-44.

If you look at the break down of the openings played, you get the following picture:

Leela white
* Sicilian Najdorf 6.Be3: 30 times
* French Defence 3.Nc3: 17 times
* Sicilian other: 3 times

Almost no variety here, with Leela playing exclusively 1.e4 and Stockfish playing the same Najdorf and French line 47 out of 50 times. Some games were identical up to move 20 or even more.

Stockfish white
* Ruy Lopez Berlin with 4.d3 and 5.Bxc6: 27 times
* Italian game 3.Bc4 Bc5: 11 times
* QGD Ragozin: 8 times
* Other 1.d4 lines: 4 times

Almost the same picture here, with Stockfish alternating between 1.e4 and 1.d4, but constantly picking the 4.d3 and 5.Bxc6 line vs the Berlin (which yields white nothing). In 46 out of 50 games they repeated the same 3 lines.
Parent - - By Labyrinth (*****) Date 2019-03-06 22:01
Still, awesome job Leela for taking down SF! Even if it's only in bookless!
Parent - By Venator (Silver) Date 2019-03-07 16:29
Suppose I would have only picked 5 lines in the superfinal and repeat them over and over again. What would people say?

No doubt Leela's result is impressive, but 100 games with the same stuff over and over again? Seriously, I have never witnessed a more boring match than this one :-).
- By Venator (Silver) Date 2019-03-02 11:23
GM Matthew Sadler's view on the TCEC 14 superfinal:
Up Topic The Rybka Lounge / Computer Chess / TCEC 14 superfinal opening PGN available

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill