Finally LCZero wins Stockfish in 100 games match

[Jouni]

NO book blitz: 50,5-49,5. Actually Stockfish lost won position by time https://tcec-chess.com/#div=bcb&game=82&season=28 . This game shows, that 7 piece tablebases have negative effect in fast games.

[Ajedrecista]

Hello Jouni:

NO book blitz: 50,5-49,5. Actually Stockfish lost won position by time https://tcec-chess.com/#div=bcb&game=82&season=28 . This game shows, that 7 piece tablebases have negative effect in fast games.

What is wrong with this game lost by time? My copy of SF 17 gives 0.00 evals at depth 57 and thousands of TB hits, nothing like those winning/losing evals at TCEC web.

-----------

Regarding this concrete game, it reminds me Japan 2-1 Spain in the last game of the Group E of the the 2022 FIFA World Cup:



It was ruled as inside the field and Japan scored the second goal, so it was a game changer. If out: drawn game, Spain first of Group E instead of second (so a different opponent team in the next round) and I guess that Japan second instead of first (I say 'I guess' because there is a chance that Germany would have been second and Japan third, when only first and second qualified to the knockout stage). Such huge changes due to few milimeters at most... and not a clear decision at all.

------------

Regarding the final result: +1=99-0 (or the possible +0=100-0 with the eval of my local copy of SF 17 and without losing by time) and we are approaching the checkers scene of some years ago of all games drawn, as I have said many times before.

Using trinomial distribution, the standard deviation is:

Code: Select all

s := sample standard deviation
µ := score
D := draw ratio
n := number of games
LOS := likelihood of superiority

n = 100
µ = 50.5/n = 0.505
D = 99/n = 0.99

s = sqrt{[µ(1 - µ) - D/4]/(n - 1)}
s = sqrt[(0.505*0.495 - 0.99/4)/99]
s = 0.005

LOS: [(µ - 0.5)/s] sigma
LOS: 1-sigma (circa 84.13%)

Not even a statistically significant result.

Regards from Spain.

Ajedrecista.

[Jouni]

My SF analysis. I have only 5 piece tablebases.

Analysis by Stockfish 17.1:
...
92.Kb4 Qb7+ 93.Ka3 Kxg5 94.Qf6+ Kg4 95.e6 Qb1 96.Qe5 Qc1+ 97.Kb4 Qc4+ 98.Ka5 Qxc5+ 99.Ka6 Qe7 100.a5 c5 101.Qxd5 c4 102.Qxc4+ Kg3 103.Qc3+
+- (6.28) Depth: 24/86 00:00:33 183mN, tb=894712

[peter]

With 30 threads of 16x3.5GHz and 32G empty hash started, 6men Syzygygs, ProbeLimit=6:

[d]5Q2/q7/2p3k1/K1PpP1P1/P7/8/8/8 w - - 0 1

Analysis by Stockfish dev-20250913-fc54d873:

92.Kb4 Kh5 93.Dh6+ Kg4 94.Dxc6 Db8+ 95.Db6 Df8 96.De6+ Kg3 97.Dxd5 Kh2 98.g6 Kg1 99.Dd4+ Kh2 100.De3 Df5 101.g7 Db1+ 102.Db3 Dc1 103.Dc3 Dh6 104.Dd3 Dc1
Tiefe: 28/89 00:02:06 4797MN, tb=23150058
Weiß steht klar auf Gewinn.
+- (76.45 ++)

And after some Forward- Backward:

5Q2/q7/2p3k1/K1PpP1P1/P7/8/8/8 w - - 0 1

Analysis by Stockfish dev-20250913-fc54d873:

92.Kb4 Db7+ 93.Ka3 Kxg5 94.Df6+ Kh5 95.Df5+ Kh4 96.e6 De7 97.Df4+ Kh3 98.De3+ Kh4 99.a5 Kg4 100.Ka2 Kf5 101.a6 Kg4 102.Kb3 Dh7 103.De2+ Kg5 104.e7 Db1+ 105.Ka3 Da1+ 106.Kb4 d4 107.De5+ Kg4 108.De4+ Kg5 109.Dxc6 Dc3+ 110.Ka4 Da1+ 111.Kb5 Db1+ 112.Ka5 Da2+ 113.Kb6 Db3+ 114.Ka7 Df7 115.Dd7 Dg6 116.Dxd4 De6 117.Dd8 Df7 118.Dd7 Dh7
Tiefe: 26/114 00:07:33 25634MN, tb=654152878
Weiß steht klar auf Gewinn.
+- (83.43)

Over Forward- Backward - lines SF gets DTM rather soon now and then into output too, even if in root position, eval drops for short time with stuck iteration in between also. Anyhow I don't see so far, how Black should deny getting one of the two plus- pawns of White converted sooner or later and 50 moves can't be an issue neither.

Here's SlowChess (because of its fine specialised mate search) after backwarding one of SF's lines, started with already full hash without Syzygy- usage after one move from root position forward 92.Kb4 Kh5:

[d]5Q2/q7/2p5/2PpP1Pk/PK6/8/8/8 w - -

Engine: SlowChess Blitz 2.9 avx20 (32768 MB)
von Jonathan Kreuzer

37/55 0:54 +M24 93.Dh6+ Kg4 94.Dxc6 Db8+ 95.Db6 Dxe5
96.c6 De1+ 97.Kb5 De2+ 98.Ka5 Kxg5
99.c7 Dc4 100.De3+ Kf6 101.Dh6+ Kf7
102.Dh7+ Ke8 103.Dh8+ Kd7 104.Dd8+ Ke6
105.c8D+ Dxc8 106.Dxc8+ (1.422.678.407) 26263

Pity with 1...(92...) Qb7+ instead of ...Kh5:


[d]5Q2/1q6/2p3k1/2PpP1P1/PK6/8/8/8 w - -

Engine: SlowChess Blitz 2.9 avx20 (32768 MB)
von Jonathan Kreuzer


34/74 0:12 +8.48 93.Ka3 Kxg5 94.Df6+ Kg4 95.e6 Db1
96.De5 Dc1+ 97.Kb4 Dd2+ 98.Dc3 Df4+
99.Ka5 Kf5 100.Dh3+ Kf6 101.Dh8+ Kxe6
102.De8+ Kf5 103.Dxc6 Dc4 104.Db5 Dc3+
105.Kb6 d4 106.a5 (387.783.772) 30486

2 plies too much to keep DTM in hash at root position

[chessica]

And you need a high-end PC for that?

[Jouni]

Actually SF has winning score without TBs:

1.Kb4 Qb7+ 2.Ka3 Kxg5 3.Qf6+ Kg4 4.e6 Qb1 5.Qe5 Qc1+ 6.Kb4 Qc4+ 7.Ka5 Qxc5+ 8.Ka6 Qc4+ 9.Kb7 Qb4+ 10.Kc7 Qe7+ 11.Kxc6 Qe8+ 12.Kxd5 Qa8+ 13.Kc5 Qc8+ 14.Kb6 Qd8+ 15.Kb7 Qe7+ 16.Ka6 Kh3 17.a5 Qf8
+- (6.42) Depth: 35/100 00:06:43 2400mN
1.Kb4 Qb7+ 2.Ka3 Kxg5 3.Qf6+ Kg4 4.e6 Qb1 5.Qe5 Qc1+ 6.Kb4 Qc4+ 7.Ka5 Qxc5+ 8.Ka6 Qc4+ 9.Kb7 Qb4+ 10.Kc7 Qe7+ 11.Kxc6 Qe8+ 12.Kxd5 Qa8+ 13.Kc5 Qc8+ 14.Kb6 Qd8+ 15.Kb7 Qe7+ 16.Ka6 Kh3 17.a5 Kg2 18.Qd5+ Kh3 19.Qh1+ Kg4 20.Qc6 Kh3
+- (6.42) Depth: 36/94 00:06:47 2423mN
1.Kb4 Qb7+ 2.Ka3 Kxg5 3.Qf6+ Kg4 4.e6 Qb1 5.Qe5 Qc1+ 6.Kb4 Qc4+ 7.Ka5 Qxc5+ 8.Ka6 Qc4+ 9.Kb7 Qb4+ 10.Kc7 Qe7+ 11.Kxc6 Qe8+ 12.Kxd5 Qa8+ 13.Kc5 Qc8+ 14.Kb6 Qd8+ 15.Kb7 Qe7+ 16.Ka6 Kh3 17.a5 Kg2 18.Qd5+ Kg3
+- (6.42) Depth: 37/114 00:06:52 2455mN

[Ajedrecista]

Hello:

I finally get winning evals. I think I set up a wrong position before, although it seemed right for me. In the middle of searching for Windows updates, I managed to run SF 17 with two threads at crazy speeds like 300 N/s then and now, so the output of depths 3, 4, 5 and so on could be watched at a slow pace. This speed is between 3 and 4 orders of magnitude below my normal speeds.

------------

Regarding the ball image, I found a diameter of the ball of 8.6 inches on the Internet. Zooming the image and measuring by myself, I think that the part being in was around 2% at most being generous, and more likely 1%. We are talking about 0.1 inches supposing that the image frame was the correct one and the ball did not roll outer than in the image. These things happen sometimes (resignation sigh).

Regards from Spain.

Ajedrecista.

[jkominek]

Hello Jouni:

NO book blitz: 50,5-49,5. Actually Stockfish lost won position by time https://tcec-chess.com/#div=bcb&game=82&season=28 . This game shows, that 7 piece tablebases have negative effect in fast games.

What is wrong with this game lost by time? My copy of SF 17 gives 0.00 evals at depth 57 and thousands of TB hits, nothing like those winning/losing evals at TCEC web.

Word in the chat discussion has it that in game 82 Stockfish spent 1.5 seconds building an inside-TB PV extension to checkmate through syzygy DTZ tablebase lookup, but there was less than that amount of time left available to complete the move. Hence white losing on time.

Stockfish 17 introduced extending the PV until checkmate. ("For mate scores, the PV shown is the best variation known to the engine at that point, while for table base wins, it follows, based on the TB, a sequence of moves that preserves the game outcome to checkmate.")

It is a user-friendly feature. But the operation evidently slides in between time checks, and so the code that ensures a move is made before exhausting the remaining time on the clock did not get activated soon enough. The furious time blitz of the endgame exposed the issue.

One would need the log file to verify this, as it is not evident in the pgn file alone. So what I am relaying belongs to the category of Things Heard On The Internet.

[peter]

Actually SF has winning score without TBs

It has, BTW, the DTM- output of Slowchess' above was without Syzygy- usage too.
Here's SF 16 with NNUE unchecked and ProbeLimit=0, after an overnight- run that was started already after quite some Forward- Backward:

[d]5Q2/q7/2p3k1/K1PpP1P1/P7/8/8/8 w - - 0 1

Analysis by Stockfish 16:

92.Kb4 Db7+ 93.Ka3 Kxg5 94.Df6+ Kg4 95.e6 Db1 96.De5 Dc1+ 97.Kb4 Dd2+ 98.Dc3 Df4+ 99.Ka5 Kf5 100.Dh3+ Kf6 101.Dh8+ Kxe6 102.De8+ Kf5 103.Dxc6 Dc4 104.Dd7+ Kf6 105.Dd6+ Kf7 106.c6 Dc3+ 107.Ka6 Dc4+ 108.Kb7 Db3+ 109.Kc8 Dxa4 110.Dc7+ Ke6 111.Kb8 Db4+ 112.Db7 Dd6+ 113.c7 Kf5 114.Db1+ Kg5 115.Kb7 Dd7 116.Db6 De7 117.Dd4 Kf5 118.Dxd5+ Kg6 119.Dd3+ Kh5 120.Df5+ Kh4 121.Kc6 De3 122.c8D Dc3+ 123.Dc5 Df3+ 124.Kb5 Db3+ 125.Db4+ Dxb4+ 126.Kxb4 Kg5 127.Kc3 Kf4 128.Kd4 Kf3 129.Da8+ Kf4 130.De4+ Kg3 131.Ke3 Kh2 132.Kf3 Kg1 133.Dg6+ Kf1 134.Db1#
Tiefe: 67/88 07:04:51 2384351MN
Weiß setzt Matt.
+- (#43)

[peter]

Having the SF 16- output- line to mate above already, Huntsman has an easier run backwarding it, after that, DTM's even one move shorter as for (again with ProbeLimit=0) new output:

5Q2/q7/2p3k1/K1PpP1P1/P7/8/8/8 w - - 0 1

Analysis by The Huntsman 1:

92.Kb4 Db7+ 93.Ka3 Kxg5 94.Df6+ Kh5 95.Df5+ Kh6 96.e6 Db8 97.Df6+ Kh7 98.Dg5 Db1 99.De3 d4 100.Dxd4 De1 101.e7 Kg6 102.Kb3 De6+ 103.Kb4 De1+ 104.Dc3 Dxe7 105.Dg3+ Kh5 106.Df3+ Kh4 107.Dh1+ Kg4 108.Dxc6 De1+ 109.Kb5 Db1+ 110.Ka6 Df1+ 111.Db5 Df8 112.Dd7+ Kh4 113.c6 Da8+ 114.Kb5 Db8+ 115.Db7 De8 116.Ka6 De2+ 117.Ka7 Da2 118.Db4+ Kh3 119.c7 Df2+ 120.Db6 Df5 121.a5 Kg2 122.Dc6+ Kh2 123.a6 Dc8 124.Db7 Df5 125.c8D Df2+ 126.Kb8 Df4+ 127.Ka8 Df2 128.Dh8+ Kg3 129.Dhg7+ Kh2 130.Dbc7+ Kh1 131.Dc1+ De1 132.Dxe1+ Kh2 133.Dh4#
Tiefe: 64/84 00:00:08 538MN
Weiß setzt Matt.
+- (#42)