zaterdag 17 oktober 2020

TCEC 19 superfinal opening PGN available

Jeroen Noomen has made his TCEC 19 superfinal book available for download. His comments:

"The TCEC 19 superfinal ended Friday, October 16th and was won by Stockfish by 54.5-45.5 vs Leela Chess Zero. In wins 18-9, with 73 draws. Just like the S18 superfinal book, the S19 book was way more risky than S16 and before. Nevertheless, the S19 book also had several balanced lines and theoretically interesting lines. Interestingly, all wins were only in (very) dubious openings. Balanced lines and slightly dubious lines had a 100% draw rate. A sign of the times?

You can download the S19 superfinal book from the following location:

http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=33689

Some interesting statistics:

In the Leela wins, the book exit evaluations by Leela were: 0.43 / 0.54 / 0.50 / 0.46 / 0.66 / 0.48 / 0.56 / 0.52 / 0.63

I.e. the lowest book exit that led to a win was 0.43 (which is quite high by Leela standards). All other white Leela games that had lower exits ended in a draw.

In the SF wins, the book exit evaluations by SF were: 1.14 / 0.99 / 0.77 / 0.63 / 0.98 / 1.39 / 1.06 / 1.19 / 1.14 / 1.20 / 0.70 / 0.66 / 1.38 / 1.05 / 0.91 / 1.27 / 1.00 / 0.85  (note: with contempt=24)

I.e. the lowest book exit that led to a win was 0.63. All other white SF games that had lower exits ended in a draw.

Conclusions:

1. All balanced lines, slightly unbalanced lines and slightly dubious lines ended in draws => the draw rate for such lines is 100%;
2. The ONLY wins in this superfinal were scored in (very) dubious lines.

zaterdag 19 september 2020

Opening selection TCEC 19 superfinal

 Here are Jeroen Noomen's comments on the opening selection for the TCEC 19 superfinal:


"Here are the most important details regarding my selection of the openings for the TCEC 19 superfinal:
  • After the TCEC 18 superfinal two bonus events were played with short, balanced openings. The result was predictable: more than 90% draws and a lot of games where evaluations stayed close to 0.00 with not much happening. Thus the conclusion is clear: to keep computer chess (and TCEC!) interesting, unbalanced openings with higher book exits are absolutely necessary.
  • The TCEC 19 superfinal book was mostly completed in April 2020. After seeing the TCEC 18 superfinal in which there were a couple of busted lines, I decided to do a thorough quick check with Stockfish and Leela on the TCEC 19 superfinal book. This led to the removal of a couple of busted lines on July 3rd, which were replaced by more proper lines. Normally I don't change books anymore when they are finished, but because of the TCEC 18 superfinal experience I decided to act and make reparations. 
  • As usual, all important openings will be played; just like in the TCEC 18 superfinal there will be a lot more risky opening lines compared to TCEC 16 and before. However, the TCEC 19 superfinal book will be less extreme than the TCEC 18 superfinal book, which was a bit too much in my opinion.
  • The chosen lines typically have a length between 1 move and 13 moves; there are four lines which are only 1 move (i.e. the first white and black move), there is one line with 14 moves of book. Of the 1 move lines one is rather topical, the other three are unbalanced.
  • A few hints: in this superfinal we will see the return of the Grünfeld Indian, an opening that always causes me headaches, as most lines are too drawish for TCEC superfinal purposes. Of course the King's gambit is present as well. 
  • Average move length is 6.44 (which is 0.6 lower compared to the TCEC 18 superfinal)
  • There are sharp lines in the set, as well as positional linesmore complex lines, lines with opposite castling ("SALC", i.e. short and long castling), dubious lines, lines with an imbalance in material and gambits; my personal favourite lines are #12, #17, #20, #28, #36 and #48 through #50.
  • As TCEC is also followed by strong grandmasters and we also see TCEC games mentioned at theoretical websites as https://www.chesspublishing.com/content/, I now also prefer to include some theoretically important lines, to see if the engines can produce interesting novelties. 
  • The complexity of a position and the book exit evaluation are the most important selection criterions. Book exits have been checked with classical Stockfish (development version, contempt = 0) on a 16 core Ryzen system and Leela Chess Zero v0.26.0, with the net that played in the TCEC 18 superfinal. As mentioned above, book exits are significantly higher compared to season 16 and before, to keep the games interesting and avoid boring draws. Nevertheless, there are also some lines with a lower book exit, especially in the theoretically interesting lines.
  • As I mentioned in my article about the TCEC 18 superfinal book, I now prefer 1-0 1-0 results over two straight forward draws. In effect the result is the same: 1-1. But two wins give more excitement. 
  • I have avoided positions that could lead to many exchanges and/or have a high draw rate in correspondence chess (I prefer draw rates of 50% or less), as well as positions with symmetrical pawn structures and positions that could lead to easy wins for the side with an advantage. The last aspect is becoming more and more difficult, with strong hardware, a long time control and engines being so strong. Of course risky lines means that the chance of 1-0 1-0 is increasing.
  • What do I expect? Frankly speaking I don't know, now that Stockfish also 'went NN'. So I am a bit cautious: 20 wins is the minimum I'd like to see. In any case: enjoy the superfinal!

ECO code distribution
ECO A: 11 lines
ECO B: 14 lines
ECO C: 11 lines
ECO D: 4 lines
ECO E10 lines

Move length distribution
1 move: 4 lines
2 moves: 1 line
3 moves: 2 lines
4 moves: 10 lines
5 moves: 6 lines
6 moves: 5 lines
7 moves: 4 lines
8 moves: 6 lines
9 moves: 3 lines
10 moves: 3 lines
>10 moves: 6 lines

donderdag 2 juli 2020

TCEC 18 superfinal opening PGN available

Jeroen Noomen's TCEC 18 superfinal opening PGN is available for download:

TCEC 18_Noomen Select

The opening set is different from all other superfinal books on several points:
  • Much higher book exits
  • Lots of very risky lines (increasing the chance of a double 1-0 or 0-1 result)
  • Most of the lines are offbeat

    Thus if you like a 'bloody' engine match, you will like this opening set!

woensdag 10 juni 2020

Opening selection TCEC 18 superfinal

Here are Jeroen Noomen's comments on the opening selection for the TCEC 18 superfinal:

"Here are the most important details regarding my selection of the openings for the TCEC 18 superfinal:
  • The book was completed in December 2019 and is the result of the (for me) disappointing TCEC 16 superfinal; in this superfinal it became clear to me that my usual opening selection method was not working anymore
  • All important openings will be played; in this superfinal, however, there will be a lot more risky opening lines compared to S16 and before
  • The chosen lines typically have a length between 1 move and 12 moves; there is one line which is only 1 move (i.e. the first white and black move), there is one line with 16 moves of book
  • I have chosen a couple of S16 user bonus lines, a line from a Karpov-Korchnoi match and in this superfinal we will also see the return of the King's gambit, which was absent in S17
  • Average move length is 7.04
  • There are sharp lines in the set, as well as positional linesmore complex lines, lines with opposite castling ("SALC", i.e. short and long castling), dubious lines, lines with an imbalance in material and gambits
  • Some lines are borrowed from previous superfinals, they can be the same, somewhat shorter, or longer, if necessary
  • The complexity of a position is one of my main selection criteria, i.e. most pieces are still on the board and there should be a lot of play left, with a reasonable chance for a decisive result. Compared to previous superfinals the book exit evaluations are mostly significantly higher. This has been checked with Stockfish (development version, contempt = 0) on a 16 core Ryzen system and Leela Chess Zero, with a December net that was suitable for my hardware
  • In each line one side typically has an advantage, so each engine gets the chance to push for a win, while in the return game it has to defend the weaker side
  • I have avoided positions that could lead to many exchanges and/or have a high draw rate, as well as positions with symmetrical pawn structures and positions that could lead to easy wins for the side with an advantage. I also have avoided 'easy', very solid, classical positions, in which there is almost no chance for the engines to go wrong and where the expected draw rate will be very high (>90%)

ECO code distribution
ECO A: 12 lines
ECO B: 13 lines
ECO C: 11 lines
ECO D: 4 lines
ECO E: 10 lines

Preferable opening results
My personal ranking of desired opening results (i.e. 2 games with each line) is as follows:
  • 2-0  This is very rare, nevertheless we had such a result in the S14, S15 and S17 superfinals
  • 1.5-0.5  One engine wins the opening minimatch
  • 1-1 with both engines winning hard fought games, that are not the result of the chosen opening
  • 1-1 with both engines drawing and the draws being very interesting and fiercely contested
As you can see, I don't object against 1-0 1-0 (or 0-1 0-1) results anymore, I now find them slightly more desirable than two hard fought draws. This is a new insight for me, as before S17 I regarded a 1-0 1-0 result as rather undesirable.

The two results I want to avoid, is 1-1 with two very onesided wins or two very boring draws, where the result is already clear shortly after the engines left book.

A personal note
My sources to select openings for the TCEC superfinals are: GM games, correspondence games, databases, chess books, theoretical websites and my personal opening experience. I prefer lines that have a rather low draw rate in correspondence chess, i.e. preferably 50% or lower.

Personally I am a big fan of original, enterprising, lively, complex and non-standard play. In my opinion lines in this category can lead to far more interesting games than in the usual solid, classical stuff. Hence my opening selection reflects that preference.

Nevertheless, my picks will also always include well known opening lines, as long as they don't have a high chance of a draw. Please note that simple, classical, open and standard opening positions just have a very high chance of a draw, especially under TCEC superfinal conditions and with the TCEC hardware. We want to avoid >90% draws, therefore I exclude such lines from my selections. I understand that might not be to everybody's liking, but I'd say: if we have a lot of exciting games, this is the way to do it!

How many wins do I expect? Well, in the S17 superfinal we had 29 wins, which was quite a lot higher than I expected. As my S18 picks are a lot riskier, I expect a minimum of 25 wins and I think that more than 30 wins are entirely possible.

So: have fun! I hope you enjoy the games and we will see lots of exciting chess and dramatic turn of events, like we had in all previous superfinals!"

dinsdag 21 april 2020

TCEC 17 superfinal opening PGN available

The opening file in PGN format used in the TCEC 17 superfinal is available for download:

TCEC 17 superfinal PGN

The book has 50 lines, which the two contestants Leela Chess Zero and Stockfish played twice: one time with white and one time with black.

Leela Chess Zero won the superfinal, congrats to the team members!

vrijdag 27 maart 2020

Opening selection TCEC 17 superfinal

Here are Jeroen Noomen's comments on the opening selection for the TCEC 17 superfinal:

"Here are the most important details regarding my selection of the openings for the TCEC 17 superfinal:
  • The book was completed in October 2019 and hasn't changed since then
  • All important openings will be played
  • The chosen lines typically have a length between 1 move and 12 moves; there are two lines which are only 1 ply (i.e. the first white move only), there is one line with 15 moves of book
  • Average move length is 6.24, which is the lowest in my TCEC superfinal career
  • There are sharp lines in the set, as well as positional linesmore complex lines, lines with opposite castling ("SALC", i.e. short and long castling), dubious lines, lines with an imbalance in material and gambits
  • Some lines are borrowed from previous superfinals, they can be the same, somewhat shorter, or longer, if necessary
  • The complexity of a position is one of my main selection criteria, i.e. most pieces are still on the board and there should be a lot of play left, with a reasonable chance for a decisive result
  • In each line one side typically has an advantage, so each engine gets the chance to push for a win, while in the return game it has to defend the weaker side
  • I have avoided positions that could lead to many exchanges and/or have a high draw rate, as well as positions with symmetrical pawn structures and positions that could lead to easy wins for the side with an advantage. I also have avoided 'easy', very solid, classical positions, in which there is almost no chance for the engines to go wrong and where the expected draw rate will be very high (>90%)
  • Note: I haven't taken into account the TCEC hardware upgrade with this opening set, this might lead to a slightly higher draw rate than usual. Still I hope that we get 15-20% wins

ECO code distribution
ECO A: 9 lines
ECO B: 15 lines
ECO C: 10 lines
ECO D: 5 lines
ECO E: 11 lines

Distribution of move length
1 ply: 2 lines
1 full move: 2 lines
2 moves: 1 line
3 moves: 5 lines
4 moves: 8 lines
5 moves: 4 lines
6 moves: 6 lines
7 moves: 6 lines
8 moves: 2 lines
9 moves: 6 lines
10 or more moves: 8 lines

A personal note
My sources to select openings for the TCEC superfinals are: GM games, correspondence games, databases, chess books, theoretical websites and my personal opening experience. I prefer lines that have a rather low draw rate in correspondence chess, i.e. preferably 50% or lower.

Personally I am a big fan of original, enterprising, lively, complex and non-standard play. In my opinion lines in this category can lead to far more interesting games than in the usual solid, classical stuff. Hence my opening selection reflects that preference.

Nevertheless, my picks will also always include well known opening lines, as long as they don't have a high chance of a draw. Please note that simple, classical, open and standard opening positions just have a very high chance of a draw, especially under TCEC superfinal conditions and with the TCEC hardware upgrade. We want to avoid >90% draws, therefore I exclude such lines from my selections. I understand that might not be to everybody's liking, but I'd say: if we have a lot of exciting games, this is the way to do it!

So: have fun! I hope you enjoy the games and we will see lots of exciting chess and dramatic turn of events, like we had in all previous superfinals!"

dinsdag 31 december 2019

Noomen LowDrawRate testset

Here is a new testset in PGN format from Jeroen Noomen:

Noomen_LowDrawRate_v1: 174 lines with a low draw probability in engine-engine matches

The 174 lines are taken from the TCEC 9 through TCEC 16 superfinal openingsets and promise the highest chance of a decisive result. Furthermore, the opening lines lead to enterprising, lively and interesting play.

With hardware becoming faster and faster and the top engines slowly but surely playing more perfectly, normal, riskless and solid opening lines will not do anymore. Especially in long time controls the chance is becoming bigger that the draw rate will creep towards 100%. The Noomen_LowDrawRate_v1 testset is designed to prevent such high draw rates.

Updates will be made available in the future. Not only TCEC superfinal lines will be added (after a superfinal has ended), but also other promising and enterprising lines.

The testset can be downloaded below:

Download Noomen_LowDrawRate_v1.pgn