24 November 2012

Waving a Yellow Flag

A few years back, in a post titled Advantage in Chess960 Start Positions Revisited, I created a table from CCRL data showing which start positions appeared to be best for White and and which best for Black. The CCRL site, linked to the right of this page under 'Resources', explains,
CCRL means "Computer Chess Rating Lists". We are a club of people inspired by watching computers play chess.

While I'm not at all 'inspired by watching computers play chess', I do accept that the CCRL findings are the single most important collection of chess960 games in existence today. I ended the 'Revisited' post by concluding that the CCRL sample at that time was probably too small to make generalizations. The club now boasts of having played 115.900 games of chess960, or an average of more than 115 games per start position, or 3-4 times more games than at the time of my first post.

How do the current CCRL statistics compare with that first analysis? Here is a table similar to my earlier table. As before, it shows the number of games played (#); the percentage of games won by White, won by Black, and drawn (%W, %B, %D); and the overall score for White (%S).

Highest Overall Score White
Highest Overall Score Black
Traditional Start
Traditional Start (K&Q switched)
534RNBKQBNR 8434.535.729.849.4

The first observation is that the 'the overall score for White (%S)' lies in a narrower range than before: 64%-42% now vs. 76%-32% earlier. The second observation is that of the top five percentages for White, only one start position is on both lists (SP868 QBBRKRNN); of the bottom five percentages for White (i.e. top five for Black), the list is completely different. The third observation is that the traditional start position (SP518 RNBQKBNR) is now in line with experience.

On the earlier list, I examined the top position for White (SP024 NBQNBRKR) in A Followup, an Error, and an Insight, and surmised (with the help of another chess960 enthusiast) that 'the dangerous start positions share the common feature of several Pawns unprotected and these can be easily attacked by the Queen and Bishop'. That is clearly the case for SP868, where the 'QBB' on the a/b/c-files rake the 'RNN' f/g/h-files. For the record, SP024 is now no.12 on the list, with a 60.9% success rate for White rather than the 76.2% in the earlier table. It's worth remembering that the CCRL requires both 'Book learning' and 'Position learning' to be 'Off for all engines'. This means that the engines don't improve by playing more games.

Now here's a big, yellow WARNING FLAG: the no.1 and the no.3 entries in the current table (SP408 RBQNBNKR & SP749 RKNBNQBR) are twins -- they have the same sequence of pieces, but in reverse order -- only the castling considerations are different. These two positions, plus the repeat position SP868, need to be examined more closely. I would also like to identify the other top (and bottom) twins to determine how random their CCRL results are. I'll follow-up these actions in future posts.


GeneM said...

The best two start setups I see in the latest CCRL data might be:

S#937 RKRBBNQN Draw=12.5% , WhiteWin=48.6%

S#494 QRNKNBBR Draw=12.8% , WhiteWin=50.3%

They both have low draw rates, and low rates for White wins.

They both avoid the undesirable bishop start on a corner square.
And they both have White's two knights start on squares of the SAME shade.


In 2012/February I posted on the CCRL blog to inquire about castling rates in chess960-FRC.
Some excellent work on data mining by CCRL enthusiasts has generated the following data about castlings in chess960-FRC:


"Chess960-FRC Setup Data is Excellent"

_________________________ START DATA POST. ______
Re: Chess960-FRC Setup Data is Excellent
by Adam Hair » Fri Mar 30, 2012 7:15 am

Here are the correct statistics (thanks to Ed Collins):

Castling Statistics:
White castled Kingside and Black castled Kingside: 26,787 (25.0%)
White castled Kingside and Black castled Queenside: 6,088 (5.7%)
White castled Kingside and Black did not castle: 12,679 (11.8%)

White castled Queenside and Black castled Kingside: 5,577 (5.2%)
White castled Queenside and Black castled Queenside: 11,422 (10.7%)
White castled Queenside and Black did not castle: 10,142 (9.5%)

White did not castle and Black castled Kingside: 10,735 (10.0%)
White did not castle and Black castled Queenside: 9,030 (8.4%)
Neither player castled: 14,640 (13.7%)

Adam's Computer Chess Pages ( http://adamsccpages.blogspot.com/ )
_________________________ END DATA POST. _______

I am disappointed that the percentage of chess960 games that include the two colors castling to OPPOSITE wings (a- and h-wings) is a lowly 10.7%. I believe the rate is about 10% in traditional chess1. Thus the increase I was hoping for in chess960 is not happening.

A higher percentage of opposite wing castlings would be desirable because such games have a lower draw rate, and there simply should be more of them for more variety.

CastleLong.com, for chess960-FRC

HarryO said...

Hi Mark
For SP408 check out this beautiful defensive setup I found in 5 minutes of study, that black adopts via the nice first move 1...c5!:

[Event "?"]
[Site "?"]
[Date "2012.11.27"]
[Round "?"]
[White "New game"]
[Black "?"]
[Result "*"]
[Variant "chess 960"]
[SetUp "1"]
[FEN "rbqnbnkr/pppppppp/8/8/8/8/PPPPPPPP/RBQNBNKR w HAha - 0 0"]
[PlyCount "12"]

{Chess 960-Position 408} 1. d4 c5! {the variations where black's queen comes out early are a wonderful study in themselves} 2. d5 e6 3. c4 f5 4. Bc3 Nf7 5. b3 Ng6 6. h4
h5 *

The thing I love about 960 is that it's ok to develop the queen early when there is no way to attack her or she immediately can go to a safe square there after without trouble.


HarryO said...

Gene, respectfully you cannot be disappointing with the lack of opposite wing castling because almost all the engines do not understand the subtleties of castling in Chess960 yet.

It's not just about whether it actually happened in fact, but also the potential for it to happen even if it didn't. There are many creative ways to protect the king and the engines do not have that type of creativity because they cannot see patterns right at the start of the game.

As well as that, both your suggested start examples SP937 and SP494 are probably polluted by the fact that almost all the engines have trouble developing the queen if they are playing black. I note that one of these contains a queen in the corner, and the other a queen hemmed in between two knights.

Cheers, Harry