SALC V5 openings and books launched

[pohl4711]

Thank you. I learn important lessons from all of your books.

I also think that short verses long castling is a very important theme.

In addition, games can become really exciting when opposite castles occur.

This is an immutable law of chess strategy, not even necessarily related to computer chess.

It's one of the first dynamics I teach and train my students in, actually.

Anyway, I love these types of "theme-based" books, and will share my thoughts more on this one when I finish work.

For me and Thomas Zipproth (Brainfish / Cerebellum) it was a huge surprise, that the results, when using SALC openings, is not distorted, when you compare it with an "normal" openingset like the HERT-set. I posted the results on my website and in the SALC V5 readme-file, but I think I should post it here again, because this is very important (it proofs, that SALC can be used for tournaments or rating-lists without distorting the results (!). And using SALC instead of any "normal" openings-set, reduces the number of draws a lot and the average game-length around 11% (!))


Using the new HERT openings-set (by Thomas Zipproth) for my Stockfish-testing was a great opportunity to compare the gamebases played with HERT (contains positions selected from the most played variations in Engine and Human tournaments) and played with my SALC openings So, here the results. Both gamebases were played with 3'+1'', singlecore, 512 MB Hash. The only difference was the opening-set (HERT / SALC)... 2x 15000 games (!)



HERT:

Code: Select all

Program                    Elo + - Games Score Av.Op. Draws

1 Stockfish 170526 bmi2 : 3346 7 7 5000 71.3 % 3171 45.6 %
2 Komodo 11.2.2 x64     : 3314 6 6 5000 66.9 % 3177 45.8 %
3 Houdini 5 pext        : 3299 6 6 5000 64.7 % 3180 48.5 %
4 Shredder 13 x64       : 3119 6 6 5000 37.8 % 3216 43.7 %
5 Fizbo 1.9 bmi2        : 3096 6 6 5000 34.4 % 3221 38.2 %
6 Andscacs 0.91b bmi2   : 3026 7 7 5000 24.9 % 3235 34.9 %

Elo-differences:
1-6: 320 (overall)

1-2: 32
2-3: 15
3-4: 180
4-5: 23
5-6: 70

Games: 15000 (finished)

average game length: +13.7% compared to SALC games (moves), +10% compared to SALC games (time)

White Wins: 5129 (34.2 %)
Black Wins: 3455 (23.0 %)
Draws: 6416 (42.8 %)



SALC V3:

Code: Select all

Program                    Elo + - Games Score Av.Op. Draws

1 Stockfish 170526 bmi2 : 3359 7 7 5000 72.7 % 3168 39.9 %
2 Komodo 11.2.2 x64     : 3327 7 7 5000 68.3 % 3175 38.5 %
3 Houdini 5 pext        : 3298 6 6 5000 64.4 % 3180 42.2 %
4 Shredder 13 x64       : 3108 6 6 5000 36.4 % 3218 35.4 %
5 Fizbo 1.9 bmi2        : 3097 7 7 5000 34.8 % 3221 31.1 %
6 Andscacs 0.91b bmi2   : 3012 7 7 5000 23.5 % 3238 27.7 %

Elo-differences:
1-6: 347 (overall)

1-2: 32
2-3: 29
3-4: 190
4-5: 11
5-6: 85

Games: 15000 (finished)

White Wins: 5476 (36.5 %)
Black Wins: 4154 (27.7 %)
Draws: 5370 (35.8 %)


Conclusions:

1) SALC lowers the draw-rate a lot (35.8%) , compared to the HERT openings-set (42.8%) - mention, that the HERT-set was optimized for a low draw-rate. Thomas Zipproth has chosen only lines, which were not too drawish. Using other "classical" openings-sets should lead to a higher draw-rate, than using HERT.
2) The order of rank is the same for all engines in both gamebases = no distorted results playing SALC.
3) The scores of the engines are not getting closer to 50%, using SALC. The Elo-differences are not getting smaller (in fact, they are getting higher! (Elo-differences rank 1 to 6: 320 Elo using HERT, but 347 Elo using SALC), which proofs, that SALC does not contain a lot of lines, which are leading to a clear advantage (and easy wins) for white or black. And bigger Elo-differences make the results statistical more reliable.
4) SALC lowers the average game duration around 10%. That means, that in the same time, +10% more games can be played, which leads to statistical more valuable results in the same time.

[pohl4711]

Hi Stefan,

Thanks for your new SALC openingbook. I always like opposite castling (I also include such positions in my TCEC superfinal sets), it adds asymmetry to the game, which is a plus considering the high draw rate between the top engines.

F.e. in games with castling on the same side, you can exchange all pawns on the other side of the board without problems. In games with opposite castling this is not possible, thus "SALC" prevents lots of pawn exchanges. Another advantage is that both kings are on the opposite side of the board, which might lead to interesting endings. Finally, you might see pawn storms and attacks from such positions.

Looking to your set, I suggest a few improvements:

1. Only use human games where both players have at least 2500 elo, IMO this improves the quality of the lines;

Of course true. Problem: Not enough games... I filtered the 25000 SALC openings out of 7 million games. With both players 2500 Elo or more only 400 (!!) of 25000 lines would stay (I tried it that moment with a quick FritzGUI search)... but all 25000 endpositions were checked by Komodo 11.2.2 with 60'' on 3 cores and 2048MB Hash

2. Cut the lines at the point where the 2nd castling has been played. This avoids unnecessary moves and/or lower quality moves after that 2nd castling;

Of course, true, too. Same problem as above: Not enough different SALC endpositions would remain!

3. Include only positions with a score between 0.30 (or 0.25) and 0.60, as positions close to 0.00 might increase the probability of a draw.

Of course, true, but the same: not enough lines would stay in for a big book... Please mention, that the biggest advantage of the SALC-project is (compared to your gambit-book for example), that the book is much bigger and 25000 lines are much stuff!

But I used smaller evaluation intervals for the 2 testsets included:
For the 500 lines set, I used an evaluation-interval of [-0.39,-0.30] and [+0.30, +0.39]
For the 5000 lines set, I used an evaluation-interval of [-0.31,-0.53] and [+0.30, +0.53]

And finally, the good news is: All Elos of all human players of the games, I used for creating SALC, are still in the PGN-file. And the Komodo evaluation is stored in the PGN-file (in the tag Annotator, you find a string "val=X", which is the evaluation of Komodo for the endposition). So, it is very easy for you or anybody else, to do filter-operations on that data and create smaller books or opening-sets out of the 25000-lines SALC_V5_hc_10m.pgn file !

Example: When using FritzGUI: Copy the PGN in an empty Fritz-database. then search for "val=? " (dont forget the <space> behind the "?") and for "val=-? " and delete those games. That deletes all games with an evaluation between [0.09, +0.09] (because the evaluation have only on digit (val=1 for example)". Then search for "val=1*" and "val=-1*" and delete those games. That will delete all games with an evaluation between [-0.19,+0.19]. And so on... Takes only some minutes and the job is done!

Regards - Stefan (SPCC)

[pohl4711]

F.e. in games with castling on the same side, you can exchange all pawns on the other side of the board without problems. In games with opposite castling this is not possible, thus "SALC" prevents lots of pawn exchanges.

Correct. And with SALC V5, I introduced new filters - half-closed and (for the smaller testsets) full-closed filters. Means, that open center-lines are not allowed at the endposition of the SALC-line. To avoid fast exchanges of rooks and queens. And (very important new feature!) possible pawn-captures in the center are not allowed in the endposition (because, a lot of (pawn)exchanges will bring the game faster into (boring, drawish) endgames(!)):

half-closed, filter 4) no pawn-capture on the center-squares (e4,d4,e5,d5) possible (means: not allowed: (white pawn on e4 and black pawn on d5) or (white pawn on d4 and black pawn on e5) - so, the position cannot get fully open after 1 or 2 played moves by the engines.

full-closed, filter 7) no pawn-capture on or into the center-squares (e4,d4,e5,d5) possible (means: not allowed: (white pawn on e4 and black pawn on d5) or (white pawn on d4 and black pawn on e5) or (white pawn on d3 and black pawn on e4) or (white pawn on e3 and black pawn on d4) or (white pawn on e5 and black pawn on d6) or (white pawn on d5 and black pawn on e6)) - so, the position cannot get fully open after 1 or 2 played moves by the engines.

[Graham Banks]

Hi Stefan,

is the PGN for the smaller book available?
If so, I can create a ChessGUI book from it.

Graham.

There is only one book included (for three GUIs) - there is no smaller (?) or bigger book. The book is based on 25000 lines and the PGN-file for it, is SALC_V5_hc_10m.pgn
Then there are two testsets in the testsets-folder, one with 500 positions and one with 5000 positions/lines. Both as PGN and EPD. It would be possible to create a book out of the 5000 lines file, but I think, that would be too small for a opening book.

If you create a book out of the SALC_V5_hc_10m.pgn file for ChessGUI, then write a PM to me or contact me over my website and then I will send you my email and you can send me the ChessGUI book. And I will add it to the SALC V5 package. That would be cool.

Regards - Stefan (SPCC)

Will do Stefan. Thanks.

[pohl4711]

Hi Stefan,

is the PGN for the smaller book available?
If so, I can create a ChessGUI book from it.

Graham.

There is only one book included (for three GUIs) - there is no smaller (?) or bigger book. The book is based on 25000 lines and the PGN-file for it, is SALC_V5_hc_10m.pgn
Then there are two testsets in the testsets-folder, one with 500 positions and one with 5000 positions/lines. Both as PGN and EPD. It would be possible to create a book out of the 5000 lines file, but I think, that would be too small for a opening book.

If you create a book out of the SALC_V5_hc_10m.pgn file for ChessGUI, then write a PM to me or contact me over my website and then I will send you my email and you can send me the ChessGUI book. And I will add it to the SALC V5 package. That would be cool.

Regards - Stefan (SPCC)

Will do Stefan. Thanks.

OK, but I need not only the book, but a jpg-picture (screenshot) of the booksettings. the Setting has to be wide (not tournament or something like that) and minimal number of games must be 1 or 0 (all openinglines are unique) and all lines are ending as draw - so the score is 50% for all lines.

I sent you a PM with my email-adress.

Stefan

[Jeroen]

Hi Stefan,

I made a SALC version using the specified conditions above, i.e.:

1. Use the 25,000 games SALC book
2. Both players at least 2500 elo
3. End eval above +0.30, or -0.30 and lower
4. Capped at the 2nd castling (if possible)

I will PM the link to you.

Kind regards, Jeroen

[pohl4711]

Hi Stefan,

I made a SALC version using the specified conditions above, i.e.:

1. Use the 25,000 games SALC book
2. Both players at least 2500 elo
3. End eval above +0.30, or -0.30 and lower
4. Capped at the 2nd castling (if possible)

I will PM the link to you.

Kind regards, Jeroen

Got it. The problem is, as I said, the size: Only 165 positions/lines were left (out of 25000)...
Thats not enough for anything.
But in the SALC V5 testset-folder I have a very, very good 500 positions openings-set, which is full-closed and in a small but not drawish evaluation-interval of [-0.39,-0.30] and [+0.30, +0.39]. Complete separate filtered and calculated with 180 seconds per endposition (instead of 60 seconds for the big half-close openings). And in a testrun all 500 positions were played twice (1000 games) and the result was overwhelming (47.7% draws and a huge Elo-spreading). That opening-set can and should be used for all serious testwork up to 1000 games in an engine-engine head-to-head.

But feel free to publish any books or openings-sets you filter out of my 25000 SALC V5 games/lines. Or use them in TCEC. No problem for me. Just add a small hint, that it came from SALC, please. Then thats 100% OK.

Regards - Stefan

[Ovyron]

3. End eval above +0.30, or -0.30 and lower

What engine provides this eval? Because in my analysis of correspondence games, for example, I find a lot of out of book positions that Stockfish claims they're equal, setting scores of <+0.10 or >-0.10, you may think those positions are drawish and should be avoided in such a book.

However, Houdini 6 with Contempt=2 tells a very different story, with evals of 0.30-0.40 and such, and when playing Hoidini's moves to Stockfish lines, Stockfish eventually agrees that lines go downhill for one side, and after backsolving scores to the root (Bookup) it turns out positions were full of life, not drawish, and would have fit in such a book, but were eaten by the horizon effect.

And sometimes both Stockfish and Houdini agree the position is near 0.00, but Komodo with Contempt=100 shows a +1.40 score (equivalent to 0.40), and after playing the variations between the engines it turns out it is right, and the drawish variations are refuted (sometimes this is easy, as it turns out the 0.00 comes from a single line that they think it's drawish, so after Komodo refutes it the score go up greatly.)

Very often a 0.00 in a position only says that the chances of both sides are equal (mainly, that black has as much chances to gain the edge as white does, so white should have done something wrong to arrive here), but doesn't say anything about how drawish the position is.

I'm not sure having some engine judge a position and excluding those that it scored close to zero is the right approach for this, considering that a stronger engine that has a more correct eval for those positions would have defeated it.

[Jeroen]

Got it. The problem is, as I said, the size: Only 165 positions/lines were left (out of 25000)... Thats not enough for anything.

For a LTC match 165 positions is more than enough, the TCEC superfinal only uses 50 positions .

Just add a small hint, that it came from SALC, please.

For my TCEC 9 and TCEC 10 superfinal books I already mentioned that I use the SALC approach too, for a number of positions. Of course I will mention SALC again for the TCEC 11 superfinal book.

[Jeroen]

I just filtered out the games from Stefan's book using the criteria I mentioned. I.e. elo both human players above 2500, eval higher than +0.30 or lower than -0.30 (I believe the eval is from Komodo), no further analysis by myself and cutting the position after the 2nd castling has been made.