SlowChess Blitz Classic 2.0

[jonkr]

Very impressive progress!

Thanks. As usual after release I think I could eventually add another 40 elo or so but further progress might be limited.

Progress +67 elo.

Thanks for the very quick testing! I was hoping some conditions might show at least +60, but +67 Elo is a better than expected first result. I'm always a bit nervous until I see at least one good result that either I've messed something up somehow what I'm seeing in my own testing won't correspond to other tests.

You spoke about trapped pieces.

Even Magnus Carlsen got trapped his queen in 2017.

Thanks, I'm glad someone other than myself checks Slow analysis sometimes.

The Magnus game you posted is very interesting, I tried analyzing some but even with some quick computer analysis figuring out this game is beyond my skill level. At the end though Magnus did make a few clear (to a computer) blunders, including blundering a mate in 27 or less.

[d]3r4/4Q1pk/p1P5/2Pq1b2/1P1p4/3R3P/5PP1/6K1 w - - 2 34
SlowChess Blitz 2.6 avx (128 MB hash, 7 threads)
[25] M27 00:42.1 321.64M Rg3 Bg6 Qh4+ Kg8 Rxg6 Rc8 Rd6 Qf5 Rd8+ Kf7 Rd7+ Kg8 Qxd4 Qb1+ Kh2 Qg6 Rd8+ Rxd8 Qxd8+ Kh7 c7 Qf5 Qh4+ Kg8 Qd4 a5 bxa5 Kh7 g3 Qf8 Qh4+ Kg6 Qe4+ Kg5 Qf4+ Qxf4 gxf4+ Kxf4 c8=Q Ke4 c6 Ke5 c7 Kd6 Qd8+ Kc5 c8=Q+ Kb4 Qb6+ Ka3 Qc3+ Ka4 Qbc6#

The trapped pieces I noticed (watching Slow's bullet games) had no safe moves, and I could tell they would probably eventually lead to a loss, so was a bit disappointing when even I could notice the trouble.

[Thomas Lagershausen]

@jonkr

I am checking my games often with Slow because it likes to attack.

The Aronian-Carlsen game is beyond over everybodys skills on this planet.

Topgrandmaster Svidler tried a little overview:



But don´t waste to much time with this game because all of us looking for Slow 2.7

[Modern Times]

At long last I've managed to put together another real release.

You can download Slow Chess 2.6 from
https://www.3dkingdoms.com/chess/slow.htm
[*]FRC is still very bad compared to standard. I think this structure will be better for FRC, but I will need to start training FRC at some point and make sure it doesn't hurt standard.

I've played an initial batch of FRC games, and this now looks comfortably stronger than 2.4 at FRC, so I'll continue with it.

[jonkr]

I've played an initial batch of FRC games, and this now looks comfortably stronger than 2.4 at FRC, so I'll continue with it.

It will be interesting to see if it maintains the performance.

I did a very quick FRC test before release and after a few hundred bullet games it dropped behind 2.4 and I abandoned the test, but I think they might be close. Just yesterday I tried running overnight FRC training games and got a net that appears to be +25 Elo in FRC. (Tested again with a full round 960x2 games.) I haven't tried it in standard to see if it's better or worse there. But I'm optimistic FRC performance can be improved for 2.7 whenever that is and it will be clearly better than 2.4

[Modern Times]

I've played an initial batch of FRC games, and this now looks comfortably stronger than 2.4 at FRC, so I'll continue with it.

It will be interesting to see if it maintains the performance.

I did a very quick FRC test before release and after a few hundred bullet games it dropped behind 2.4 and I abandoned the test, but I think they might be close. Just yesterday I tried running overnight FRC training games and got a net that appears to be +25 Elo in FRC. (Tested again with a full round 960x2 games.) I haven't tried it in standard to see if it's better or worse there. But I'm optimistic FRC performance can be improved for 2.7 whenever that is and it will be clearly better than 2.4

Completed now

http://ccrl.chessdom.com/ccrl/404FRC/

[matejst]

Anyway, Jonathan, SlowChess makes a lot of progress following its own path, and I hope you will continue to develop it. Although the nets you use are much smaller, I feel that the difference between SC and SF is mainly a question of speed, while the evaluation is closer. Among others, I tried it in closed positions of the French where engines usually overestimate white's spacial advantage and it was surprisingly good. I feel that there is quite a few elos in a bit bigger nets and speed improvements.

[jonkr]

Completed now

http://ccrl.chessdom.com/ccrl/404FRC/

Thanks for testing, I'm glad it's finally above 2.4 in FRC. My net that tested +25 elo in FRC tested about -6 elo in standard. I ran some more training and have a net that tests +35 elo in FRC to 2.6. Now that it's definitely stronger I will try to spend more time FRC testing when I'm working on 2.7. Hopefully one net can be good for both, but if needed I could have Slow choose a different general net for the FRC variation, my general nets are only 500Kb.

Anyway, Jonathan, SlowChess makes a lot of progress following its own path, and I hope you will continue to develop it. Although the nets you use are much smaller, I feel that the difference between SC and SF is mainly a question of speed, while the evaluation is closer. Among others, I tried it in closed positions of the French where engines usually overestimate white's spacial advantage and it was surprisingly good. I feel that there is quite a few elos in a bit bigger nets and speed improvements.

I think maybe I can get around 10% speed improvement doing an optimization pass on what I have. The search is probably always going to be slower than stockfish (probably lose at least 20% nps there) because it does more checking for tactical/interesting moves nearer to the leaves of the tree. This was +30 elo early in development, and +12 elo last time I tested it, but that was almost a year ago.
NNUE packing so many weights in an incremental first layer is definitely a big part of it's weight count to NPS ratio. I plan to experiment more with variations on my structure, but probably no big changes unless it stops improving.

[Werner]

Hi,
is it possible version 2.6 (and others) do not use SYZYGY bases from SSD although the path is set ?

[jonkr]

Hi,
is it possible version 2.6 (and others) do not use SYZYGY bases from SSD although the path is set ?

Syzygy bases have been working on my computer and in the online tournaments I've seen, though it's always possible there's some issue.

If you're using UCI Slow does send syzygy tablebase node stats, that's the best way to check. Just from eval it might not be as obvious since Slow will still search tablebase positions to try to find mate shortest mates, and in case opponent doesn't have tablebases still play more pressing moves. Also my tablebase win score isn't nearly as high as other engines just it will be 10 or more depending on eval of position.

[jonkr]

You can download Slow Chess 2.7 from the webpage :
https://www.3dkingdoms.com/chess/slow.htm

• Standard chess is improved by about +35 Elo compared to 2.6.(10000 bullet game gauntlet against similar strength engines, 8-move book.)
• FRC play is improved by +115 Elo (in a 3840-game bullet match against 2.6 )
• I restarted a new opening book so it should have fewer bad lines, but still not a competitive book.
• As usual, the only net compatible with 2.7 is the net that comes with it. (Eventually this may change but I'm always fiddling with my net structures and it isn't based on anything else.)

If you want to read in more detail about changes this version, check this thread : http://talkchess.com/forum3/viewtopic.p ... 92#p902881