Looking for a *.pgn tool that can do the following ... can be very interesting!!

[Frank Quisinsky]

Hi Ferdy,

nice to hear from you.

Let me look in the very late afternoon today.
It's friday, a short working day and weekend ...

I think the eval have to do with a mate?!
I will search the game.

Have to make the updates for round 8 in the afternoon at first.

BTW:
I need more short games and hope the tourney are able to give me what I need. I am working a longer time on a new test-set but I made a break for the time from one year after some private changes in the last year.

After all my experiments with openings and all the super strong available engines today I am very sure that such opening positions, produced most short games, is the best possibility to test super-strong engines today. To find out / select out most interesting balanced lines isn't easy. Means to create a test-set with 500 of such positions.

This new tool and the still-running tourney helps a lot to find out which engines can help me later for ... test the own test-set.

To evaluate with winnig probability according to engine itself ...
If possible the result will be that more as 80 ECO codes will not play by stronger engines itself. All the ECO codes are to drawish or engines not like to play in NN times.

All isn't easy.

In short game databases from super GMs are already after a few moves blunders and helps a little bit only. Most interesting are the short games with attacking moves against the king in non castling positions, or better in the middle of the board. Velvet helps here a lot. I believe the secret is to optimate such strong attacking engines like Velvet with FRC positions.

Best
Frank ... the master of desaster!

[Frank Quisinsky]

Example:

[pgn][Event "40 Moves in 20 min"] [Site "fcp-tourney-2024, WASP-3"] [Date "2023.10.25"] [Round "8.22"] [White "Velvet 5.3.0 NN"] [Black "StockDory Starfish 0.1 NN"] [Result "1-0"] 1. d4 {book 0s} d5 {book 0s} 2. c4 {book 0s} e6 {book 0s} 3. Nc3 {book 0s} Nf6 {book 0s} 4. Nf3 {book 0s} Nbd7 {book 0s} 5. Qc2 {book 0s} Be7 {book 0s} 6. cxd5 {book 0s} exd5 {book 0s} 7. Bg5 {book 0s} c6 {book 0s} 8. e3 {book 0s} Nf8 {book 0s} 9. Bd3 {book 0s} Bg4 {book 0s} 10. Ne5 {+0.42/22 22s} Bh5 {+0.76/28 37s} 11. O-O {+0.44/23 28s} Bg6 {+0.71/28 44s} 12. f4 {+0.52/24 29s} Bxd3 {+0.27/30 32s} 13. Qxd3 {+0.46/25 39s} N8d7 {+0.35/25 54s (Ne6)} 14. Rf3 {+0.92/22 33s} Ng8 {+0.27/27 51s (h6)} 15. e4 {+1.03/25 36s} f6 {+1.30/28 40s (Nxe5)} 16. exd5 {+2.27/23 31s} fxe5 {+0.85/27 47s (cxd5)} 17. fxe5 {+2.33/26 1:12m} Bxg5 {+1.34/27 40s} 18. Ne4 {+2.64/27 1:19m} Qb6 {+1.77/25 42s (Be7)} 19. Nd6+ {+5.29/23 1:52m} Ke7 {+2.40/25 22s (Kd8)} 20. Rf7+ {+7.87/23 28s} Kd8 {+1.05/18 0s} 21. Nc4 {+8.48/25 1:31m} Qb4 {+5.77/29 39s (Qa6)} 22. a3 {+9.46/25 1:09m} Qa4 {+6.49/30 19s (Qe7)} 23. b3 {+10.61/22 20s} Qa6 {+6.20/26 40s (Nh6)} 24. Qg3 {+15.21/24 1:01m} Be7 {+10.61/25 35s (cxd5)} 25. d6 {+19.81/21 27s} g5 {+13.12/25 40s (b5)} 26. e6 {+26.93/22 22s} Bf6 {+20.28/24 39s (c5)} 27. Rxd7+ {+49.38/21 24s} Ke8 {+24.27/25 23s} 28. Rf7 {+M15/24 14s} Nh6 {+25.21/26 42s (c5)} 29. Qf3 {+M9/26 17s} Bxd4+ {+41.50/26 42s} 30. Kh1 {+M8/31 26s} g4 {+47.15/28 42s (Qb5)} 31. Qf4 {+M6/42 27s} Rg8 {+64.03/32 42s (Qb5)} 32. Rf1 {+M4/69 31s} Bf2 {+M6/36 42s (Nf5)} 33. Qf6 {+M2/127 0s} Rg7 {+M1/64 0s (Nxf7)} 34. Qe7# {+M1/127 0s} 1-0 [/pgn]

Much stronger engines like Velvet have no idea what Velvet like to play here ...
Where I am working ...

[Frank Quisinsky]

After my first experiments with this game ... first moves of this line is great for a test-set position.

After all, often I am thinking if my test-set is ready for take-off ... non attacker will be have a very big problem with such a test-set I like to have or I like to create. And later in a tourney with my test-set positions the final results from non-attacking but stronger engines will be to bad and not real.



On the other hand ... I will see exactly this one!
Engines in a tourney with a test-set of balance positions ... where the probably is high that a short-game can be produced. End of the day such a test-set produced lesser draws.

[Frank Quisinsky]

-387 comes from Caissa with -Mate.

Often is to find (book) end of the game.
But this problem produced the Shredder GUI.
I sent Stefan this and different other mistakes I found in the last years ... I believe I do that for 2 or 3 years ago.

Many things are not right.
Mate distance is a big topic or engines are thinking minutes for mate in 1, 2, 3 and so on.
Engines without endgame bases are not able to give mate with KNB-K
And with endgame bases different engines have bigger problems with mate-distance.
A real loss of time is when after 49 moves engines like to play a pawn move in clear draw positions and the games goes over 200 moves. All the secret and not adjustable contempt settings ...

Or the computing time for KBp-K ... wrong bishop endgames in clearly draw position.
Engines don't know that KNN can't win ...
This list is long.

And all of this and much more you can find in the database from TOP-50 engines!

Anyway, many of such problems will never change!
Most like to play 1+1 games with resign=on ... not my world of computer chess.
Computer chess programs on first positions in the World should play with more intelligence or humans lost fast interest on it if they are looking more in detail. End of the day 99% of humans like to analyze here games with Stockfish because they are thinking all the others are buggy. But we have so many nice and perfect working engines.

If one or two programmers from the participating engines are looking in the games in detail and try to fix such problems ...
A very big success because the move-average will be reduced.

Often, I lose a bit interest in engine testing for such reasons.
Move-average is the bigger topic as the high draw stats today.

The other side of the medal ... are the great playing-strength engines can produce today.

PS:
Results after round 8 (with original *.pgn from Shredder GUI) can be found under:

v8.0.0
https://www.amateurschach.de/download/_ ... y-2024.zip

[Frank Quisinsky]

Ed, you wrote:
"High evals come from NNUE training, you can lower them with the Evalcorrect option, but then you will lose some elo."

If so, why the eval-average from so many other engines are clearly lesser?

Same with not adjustable contempt:
Today many programmers set a higher contempt, not adjustable with the result: Horror endgames
They are thinking I can make few Elo more ... should be history with the average of strength from TOP-50.
Helps only vs. engines not optimated for endgames or clearly weeker engines.

The secret is to force very fast a draw if a draw is really in the near.
The move-average from your programmer friend Chris is fantastic. Chris is an engine-doctor.
OK, from Rebel clearly better in the last versions, after all I can see.

Same with NNUE training and all the high evals.
8 moves later the eval goes down ... nice to see with your new tool.

Much more interesting is to do the following ...
Delete all the draws before you start your tool.



Best
Frank

Rebel is clearly improved in style since the last versions. The late mid-game is very strong, so the attacking moves comes after stats a bit later if I compare with Velvet, Texel, Wasp, Revenge and SlowChess. So the playing-style you produced is really interesting and your engine have an very own face.

[Rebel]

Ed, you wrote:
"High evals come from NNUE training, you can lower them with the Evalcorrect option, but then you will lose some elo."

If so, why the eval-average from so many other engines are clearly lesser?

Impossible to say, so much variety in NNUE programming. It's like exploring a dark cave with only a flashlight looking for hidden treasures by trial on error.

Same with not adjustable contempt:
Today many programmers set a higher contempt, not adjustable with the result: Horror endgames
They are thinking I can make few Elo more ... should be history with the average of strength from TOP-50.
Helps only vs. engines not optimated for endgames or clearly weeker engines.

The secret is to force very fast a draw if a draw is really in the near.
The move-average from your programmer friend Chris is fantastic. Chris is an engine-doctor.
OK, from Rebel clearly better in the last versions, after all I can see.

Same with NNUE training and all the high evals.
8 moves later the eval goes down ... nice to see with your new tool.

Much more interesting is to do the following ...
Delete all the draws before you start your tool.



Best
Frank

Rebel is clearly improved in style since the last versions. The late mid-game is very strong, so the attacking moves comes after stats a bit later if I compare with Velvet, Texel, Wasp, Revenge and SlowChess. So the playing-style you produced is really interesting and your engine have an very own face.

Glad I could help.

[Frank Quisinsky]

Ed, you wrote:
It's like exploring a dark cave with only a flashlight looking for hidden treasures by trial on error.

If you will do that ...
I would love to be on the team exploring the cave.
We find the treasure of the Knights Templar, buy more hardware, and we employ people to test the engines and enjoy the sun on the beach with Battida / pineapple juice and many nice ladys ... loud thinking only.

A long way ... for the moment we have to fighting.
Maybe you have a part of the Treasure map ... 37 moves to mate!

Hm ...

[pgn][Event "40 Moves in 20 min"] [Site "fcp-tourney-2024, WASP-3"] [Date "2023.10.29"] [Round "9.24"] [White "Rebel EAS NN"] [Black "Devre 4.0 NN"] [Result "1-0"] 1. Nf3 {book 0s} d5 {book 0s} 2. c4 {book 0s} d4 {book 0s} 3. b4 {book 0s} c5 {book 0s} 4. e3 {book 0s} dxe3 {book 0s} 5. fxe3 {book 0s} cxb4 {book 0s} 6. d4 {book 0s} g6 {book 0s} 7. a3 {+0.47/25 24s} bxa3 {+0.26/24 55s} 8. Nc3 {+0.26/28 32s} Bh6 {+0.39/23 32s} 9. Qa4+ {+0.71/25 41s} Bd7 {+0.48/26 1:05m} 10. Qxa3 {+0.38/24 34s} Nf6 {+0.43/25 43s} 11. Be2 {+0.79/26 27s} O-O {+0.39/24 23s} 12. Rb1 {+0.56/27 24s} Nc6 {+0.44/25 1:05m} 13. Rxb7 {+0.89/27 41s} Bc8 {+0.28/25 49s} 14. Rb5 {+0.95/26 52s} a5 {+0.22/24 1:01m} 15. O-O {+1.13/24 24s} Bd7 {+0.27/21 17s} 16. e4 {+1.17/24 30s} Bxc1 {0.00/29 52s} 17. Qxc1 {+1.41/26 25s} Nb4 {0.00/26 12s} 18. Rc5 {+1.75/26 22s} Bc6 {0.00/24 1:00m} 19. Qh6 {+2.50/25 24s} Ng4 {+1.11/22 46s} 20. Qh4 {+3.15/27 30s} Ne3 {+1.40/23 25s} 21. Nd5 {+4.40/23 26s} Bxd5 {+2.32/23 53s} 22. Ng5 {+9.77/21 27s} h5 {+3.37/24 52s} 23. Bxh5 {+9.50/24 28s} Kg7 {+4.22/22 28s} 24. Bxg6 {+10.68/24 2:32m} Rh8 {+3.29/21 9s} 25. Bh5 {+10.17/22 37s} Nxf1 {+4.33/20 6s} 26. Qg4 {+10.74/20 5s} Qg8 {+2.66/21 54s} 27. Ne6+ {+M11/55 14s} Kh7 {+46.50/17 36s} 28. Qf5+ {+M10/63 20s} Kh6 {+M9/18 33s} 29. Qf4+ {+M9/69 17s} Kh7 {+M8/19 23s} 30. Bxf7 {+M8/75 21s} Qxg2+ {+M7/19 25s} 31. Kxg2 {+M7/74 0s} Rhg8+ {+M6/19 44s} 32. Kh1 {+M6/82 8s} Ng3+ {+M5/22 37s} 33. hxg3 {+M5/116 5s} Rg6 {+M4/22 18s} 34. Bxg6+ {+M4/116 2s} Kxg6 {+M3/25 41s} 35. Qf5+ {+M3/116 0s} Kh6 {+M2/29 12s} 36. Qg5+ {+M2/116 0s} Kh7 {+M1/100 0s} 37. Qg7# {+M1/116 0s} 1-0[/pgn]

[Peter Berger]

Rebel is clearly improved in style since the last versions. The late mid-game is very strong, so the attacking moves comes after stats a bit later if I compare with Velvet, Texel, Wasp, Revenge and SlowChess. So the playing-style you produced is really interesting and your engine have an very own face.

How do you actually improve in style? There even seems to be some kind of consensus when it is about this, judging by tools that are supposed to give data on who is the coolest style-wyse.

I completely fail to understand this whole concept. Isn't this a matter of taste? There was a time when there was just one engine that was able to do some things, like Junior8*s or 9's positional pawn sacrifices - that no other engine could do. But these days?

For reasons unknown I personally hate the way Komodo plays chess. Others seem to think it is plenty cool. But who is supposed to be the judge?

[Frank Quisinsky]

Hi Peter,

different things can be influenced and in most of cases to see with statistic only. Today much more difficult to find out something and more often personal impressions turn out to be wrong. But I think "playing-styles" was all the time in computer chess the most interesting point. To compare engines with the strength in ELO is OK ... but to compare engines in style is much more exciting.

I am not able to evaluate a style end of mid-game. Exactly the point where the TOP engines made the most improvements. After the opening moves we have a lot very good written theory books and in combination with statistics we can find out interesting features.

Maybe in 5 years KI can produced a copy of the style from each grandmaster.
Normaly KI should need only the games for behaviorel patterns.

Yes ... the storys with the style around Junior.
We know the strengths and waiting during watching the games of it.

The Kg1-h1 moves in chess computer times ...
The a3 and h3 moves from TheKing if all pieces are on good positions.
Early Hiarcs versions starting with incredible moves the king attack and druing the combination Hiarcs like to play suddenly on queen-site.
And all the 1000 impressions more ...
In the past all are more easy as today!

The best move for an engine must not be the best move for humans.
Often I am thinking engines should after opening moves not exchange to fast the pieces like bishops and knights. But engines find a small advantage with an exanchge and the game is for humans boring. So what is more important? To create a new version 50 Elo stronger or to create a tactical monster.

Engines should try more with storming pawns to make the position more complicated (of course the human view). Often the second best move or the third best move is more interesting for a more spectacular playing style.

Perhaps random that the first phase of a chess game is for me most important and interesting. Seems to be clear that I try to looking more in details. The end of the game is for me today not longer important. Engines today are able to find almost everything. Today interesting only is the first playing-phase and I am thinking ... for the next 10 years will be this on not changed.

Best
Frank

[Peter Berger]

Yes ... the storys with the style around Junior.
We know the strengths and waiting during watching the games of it.

Junior 8 was ahead of its time and could do some things other engines couldn't. I looked up the "Computer Chess Club Archives" and the "Junior-Crafty hardware user experiment" where it played an even match against Crafty on 10 times slower hardware. The games were huge fun to follow for me.

I agree with some other things you posted. I wonder how computers learned to stop this silly Kg1-h1, Kg8-h8 stuff. You don't see this with modern chess engines.