Not logged inRybka Chess Community Forum
Up Topic The Rybka Lounge / Computer Chess / New: Noomen 2-move Testsuite
- - By Venator (Silver) Date 2017-10-22 07:56 Upvotes 3
Hi all,

Inspired by TCEC season 10, stage 1, I have made a new testset: Noomen 2-move Testsuite.

200 lines with only 2 book moves (4 ply), for engine-engine testing. See the attachment below.

Enjoy testing!

Attachment: Noomen2-moveTestsuite.pgn (42k)
Parent - - By user923005 (****) Date 2017-10-23 18:22
Thanks for your efforts, as always.
Parent - - By Venator (Silver) Date 2017-10-24 12:06
Thanks, Dann. Always a pleasure to release something that can be useful to other people in the computer chess community.
Parent - By user923005 (****) Date 2017-10-24 19:39
Epd version (with interior nodes analyzed as well) attached
Attachment: jn-2move.7z - Analyzed Epd for Jeroen's 2 move book (including interior book nodes) (33k)
Parent - By Venator (Silver) Date 2017-10-30 17:08 Upvotes 2
Also added Noomen 3-move Testsuite: 500 lines with only 3 book moves.

Download link:
Parent - - By bob (Gold) Date 2019-11-16 16:16
Something you could do that would help MANY would be to rather than give shallow openings, instead give the FEN for what you consider a representative group of positions that are not instantly won or lost, but are thematic to various opening systems (IE Sicilian, Ruy, etc.)  My experience shows just how hard this is to do using a random sort of choosing algorithm.
Parent - - By MrKris (***) Date 2019-11-17 06:09
Shallow openings are a big fad for the last few years.

Here is his older thematic suites.

His World Champion and TCEC Superfinal are close to what you are referring to, I think.

I much prefer .pgn's because FEN/EPD's can not be sorted by ECO, move length etc.
Parent - - By bob (Gold) Date 2019-11-19 03:54 Upvotes 1
I also like PGN.  But FEN offers a simplicity when dealing with lots of positions.  IE 8000 positions is 8000 lines.  Yes, you lose move history and repetition information, but it does make for easier testing.  And some of the testing being done (not by me) only works with FEN.  And yes, you can encode FEN into PGN, but rather than that I would prefer normal start position PGN since adding the FEN tag still results in several lines per position.

The real problem is the number of different positions required.  And how balanced they are across various openings, etc.  Not a particularly easy task to automate (at least for me).  IE a simple idea like "generate a random number between 1 and 20 and take the position with that move number from the first game.  If it is the move after BxN, where NxB is the natural reply, it is not so good for testing a particular position (say the "simtest" approach that has been described in the past).  I'm not even sure this idea is practical unless a team of people invest thousands of hours in the selection process.

Probably one of those "ideas that sounds good until the complexity is analyzed carefully, then it looks like crap..."
Parent - By MrKris (***) Date 2019-11-27 01:05 Edited 2019-11-27 01:16
Yes, various very interesting, good ideas seem to be impossible to automate.

Someday I would like to try something like this:

- From a human database use polyglot with mainly a percentage found limit, each position in the resulting bin book seen at least x times, say 100 if 100,000 game database.

- If say something like 1.f4 was seen x times then in only a few plies the branches would fork to less than x times each, leaving just one (or a few) short main 'trunks'.

- The most main lines (Najdorf etc.) would stay above x times for many moves, so would all be in the resulting bin.

- A high ply limit would catch all the mains, if too many variations then x would be too low.

My idea would be to get the most important for humans variations, no engine 'tuning'.
A possible problem is that x would be better for some main openings than others (an automation aspect).

The old standby 'general' .bin's included with Scid and Scid vs. PC: gm2600.bin and Elo2400 look like they might have been made with a roughly similar method, but I have not used polyglot enough to say (and there is performance.bin and probably others).

- - - -

I've just now had a chance to look a Noomen's new 4-move books:
Although short lines, they do seem "thematic", they come sorted by the opening name (Ruy Lopez, Sicilian etc.) in the white player field.
- - By Venator (Silver) Date 2017-12-17 08:21 Upvotes 1
The Noomen 2-move Testsuite and Noomen 3-move Testsuite are now also available as 3-move.ctg ChessBase book:
Parent - - By gsgs (***) Date 2018-02-13 02:00 Edited 2018-02-13 02:44
I get a google warning on that download.
google were trying to make a user-profile of me and giving it
to advertisers and whether I wanted to block that.
[I never had that before, afair)

what were the criteria to select these openings ?


thanks also to user923005 for the extensive commentation and analysis, 325 lines = fens = openings,
of which

10-17 entries per line, separated by ";" ... (trying to figure this out)
Parent - - By user923005 (****) Date 2018-02-13 22:36
Most of the EPD simply follows the EPD part of the PGN standard:
(See section 16.2).

However, I add my own fields:
cce is computed centipawn evaluation which is based on wins/losses and draws.  I did a parabolic least squares fit to approximate the computer scores.
white_wins black_wins and draws are the number of games with particular outcomes in my filtered game database (not all games have outcomes, some are *)
I use pm in a non-standard way.  The PGN Standard says that the pm is just the first node of the pv, but I consider that both redundant and useless.  I use pv as the collection of the most likely moves to be played.  If there is a number in curly brackets after the suggested move, that number is how often the move was actually played.

So, given this:
rnbqkbnr/pp1ppppp/8/2p5/2P5/2N5/PP1PPPPP/R1BQKBNR b KQkq - acd 36; acs 148; bm g6; c3 "Nf6"; cce -26; ce 0; id "es-dc-neutral.0524";
pm Nc6 {9274} Nf6 {4553} g6 {3298} b6 {654} e6 {588} e5 {400} d6 {35} d5 {4} a6 {3} Na6 {2} f5 {2} b5 {1} h6 {1}; pv g6 g3 Bg7 Bg2 Nc6 e3 e6 Nge2 Nge7 O-O O-O d4 cxd4 exd4 d5 cxd5 Nxd5 Re1 Qb6 Na4 Qd8 Nac3; white_wins 6593; black_wins 5528; draws 6011; Opening English Opening: Symmetrical Variation. Normal Variation 1.c4 c5 2.Nc3; CaxtonID: 523; ECO: A34;

Postition, side to move, castle rights and ep status:
rnbqkbnr/pp1ppppp/8/2p5/2P5/2N5/PP1PPPPP/R1BQKBNR b KQkq -

Other fields are described here:
acd 36; means the depth of the analysis was 36 plies
acs 148; means that the position was analyzed for 148 seconds
bm g6; means that g6 appears to be better than other moves.
c3 "Nf6" means that I was studying Nf6 at some point or the position appeared in a famous game and Nf6 was played.  I use that comment field to hold suggested moves for study.
cce -26; means that in actual game outcomes, black is behind by about 1/4 pawn
ce 0 means the computer says the position is even
id "es-dc-neutral.0524"; means it is contained in the test set es-dc-neutral (a collection of severely balanced openings)
pm Nc6 {9274} Nf6 {4553} g6 {3298} b6 {654} e6 {588} e5 {400} d6 {35} d5 {4} a6 {3} Na6 {2} f5 {2} b5 {1} h6 {1};  means that Nc6 is the most likely move, followed by Nf6, and g6
pv g6 g3 Bg7 Bg2 Nc6 e3 e6 Nge2 Nge7 O-O O-O d4 cxd4 exd4 d5 cxd5 Nxd5 Re1 Qb6 Na4 Qd8 Nac3; is the computer plan for this position
white_wins 6593; black_wins 5528; draws 6011 are simply the game outcomes from my database for this position
Opening English Opening: Symmetrical Variation. Normal Variation 1.c4 c5 2.Nc3; CaxtonID: 523; ECO: A34; is the name and description of the opening
Parent - - By gsgs (***) Date 2018-03-03 08:08
is it somehow automated , can others use it ?
with reference, using their database
Parent - By user923005 (****) Date 2018-03-03 08:29
I have computerized processes running analysis on several machines.
Anyone is welcome to use it.
It is just information.

Nobody owns truth
Up Topic The Rybka Lounge / Computer Chess / New: Noomen 2-move Testsuite

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill