Definitive Analysis of 11-Man Ballot

null

The eleven men above certainly aren't playing 11-man ballot or any other variety of checkers, though we'll bet many of them did surely play checkers in a slightly less risky environment.

null

But today, our subject is indeed the 11-man variant of checkers. Created around 1907 by champion player Newell W. Banks, it provided an inventive alternative to the common "go as you please" (GAYP) style, in the hopes of lending the game fresh variety and reducing the number of draws.

In GAYP, of course, there are no restrictions on opening moves. Players indeed go where and as they please. The relatively short-lived "2 move restriction" style, in which the first move for both Black and White is chosen by lot, was played for a while in the early 20th century; and "3-move restriction" in which the 3-move opening sequence Black-White-Black is randomly selected, came just a bit later and is current today. (GAYP is also still played at the highest levels of the game.)

The 11-move ballot variant got some play and some attention but never really caught on, at least not until recently, and while 3-move restriction has been extensively analyzed, 11-move ballot has remained mostly new territory.

For a full description of 11-man ballot, you can see this link, but the basic idea is that one piece chosen at random is taken off the board on each side, and then the first Black and first White move are also chosen at random. There are some further rules and limitations, but that's the main idea. There are 2,500 independent ballots in this variation! Some are known to be dead lost (and are not played), many are clearly playable, and some are open to speculation.

null
Ed Gilbert
Photo Credit: Carol Gilbert

Until now. For Ed Gilbert and his mighty KingsRow computer engine have been very busy, and Mr. Gilbert has made yet another contribution to the game: he has analyzed and classified every last one of the 2,500 ballots, labeling each as won, lost or drawn for Black or White, and further, calculating a difficulty factor for all of them. This information is handily summarized in a series of tables (see links at the end of this article).

Ed has also prepared a special 11-man ballot computer opening book containing 1.1 million positions, which is now distributed with the latest download of KingsRow. He has made all of this material available free of charge on his website (again please see links at the end of this article).

11-man ballot will never be the same again.

What follows is a fairly detailed discussion of Ed's work. But first, to illustrate the depth of Ed's discoveries, take a look at the following position.

11-man Ballot No. 2493
Remove 6 and 21; 12-16 22-18

WHITE
null
BLACK
Black to Play and Draw

B:W18,23,24,25,26,27,28,29,30,31,32:B1,2,3,4,5,7,8,9,10,11,16

This, gentle reader, is a draw. Yes, it is, requiring no less than 20 star moves on the part of Black. Ed ranks this as the toughest of all the 11-man ballots and we call it the toughest checker problem ever published.

Care to try it? You can access our no-spoilers animation here. Try to guess each Black star move and the strongest reply by White before you click the arrow to show the next move. If you can solve it, you're superhuman, but it's fun to try and will give you a real workout. If you'd rather just play through on a board or on your computer, you can click on Read More to see the list of moves.

And now let's let Ed describe his work in his own words, through a series of emails sent to us over the past few months. A bit earlier on in the process, Ed said this:

null

"I have finished the analysis work I've been doing on 11-man ballots.

"I have found a few more 11-man lost ballots. There are now 248 lost ballots identified by the opening book generator. This may be the exact right number, or perhaps I'm off by 1, I can't be absolutely sure, but it's been over a month of book expansion since the last lost ballot was identified, so I feel that the analysis is quite solid now. I read on the NC Checkers site that when they play an 11-man championship match, after selecting a ballot at random, the 2 players sometimes spend some time trying to determine if the ballot is playable or perhaps a loss. Maybe this analysis will at least be useful to simplify that part of the matches."


A little later on, Ed sent this update, wherein an additional ballot (the one diagrammed above) was found to be a draw rather than a white win, reducing the number of lost ballots to 247:

"Using a combination of engine matches and additional book expansion, I found that one of the 11-man openings that I had classified as a white win is actually a draw. It is ballot #2493, and a very difficult draw to be sure. When I updated the rating tables, this ballot went to the very top of the difficulty table with the highest difficulty figure of all the drawn ballots. It appears that for the first 20 moves, there is only one drawing move for black (when white plays the strongest attack)."

null

Finally, here are the details about how Ed computed difficulty rankings. To determine a reference for relative difficulty, Ed first re-analyzed the 3-move ballots and correlated various measures with the difficulty ratings assigned by Grandmaster Richard Pask and published on the ACF website (see link at bottom of article).

Warning: The following material is not for the faint of heart or the non-mathematical! (Note: The text below contains a small post-publication correction Ed sent us based on correspondence with Martin Fierz.)

null

"Each position along the PV is visited, and a 5-second KingsRow search is done with KingsRow configured to use only an 8-piece endgame database. The sum of those search scores is the value used.

"By the way, I found that if I did the searches with a 10-piece database, the results did not correlate nearly as well as they do with 8-piece. Too many searches return 'database draw' using the 10-piece database, so too much heuristic information is lost that way. I should also mention that database draw search scores of +/-1 and repetition draw scores of +/-3 were given the value 0 during the PV sum process.

"'Match score difficulty' is the second rating value that is an independent input into the model. To get this value, a number of engine match games were played for each 11-man ballot. To be specific, each ballot was played for 96 games, KingsRow vs KingsRow, opening books off, 8-piece database, and blitz time controls of 0.1 seconds/move."

"This component of the model is based on a suggestion from Martin Fierz. I will paste exactly what he wrote to me, because I liked his reasoning and found it slightly amusing. He wrote, 'My feeling is that at 0.1s the engines are playing very well according to human standards, but not yet inhumanly well, so that this metric kind of approaches what strong human players would do when playing these ballots.'

"To turn the Black wins ('bwins'), White wins ('wwins'), and games into a ranking score, I used this logic:

if (bwins >= wwins) 
rank = bwins / games
else    
rank = -wwins / games

"I had previously tried other logic, like 'rank = (bwins + wwins) / games', but that metric by itself did not correlate as well with Richard Pask's ratings (for 3-move ballots). To put this into words, it seems that the Pask ratings are based primarily on his perception of how difficult it is for the weak side to play these ballots, and does not take into account that perhaps there are significant opportunities for the weak side to actually score a win. It didn't happen often, but you can see there are a few ballots where the weak side scored almost as many wins as the strong side.

"So we have these two rating values, PV sum and match score, and these were combined and a best fit to Pask's scaled ratings was found through linear regression. If you plot a best fit of each of these ratings to Pask's separately, you find that PV sum by itself has an R-squared value of 0.608.

"When the match score is used by itself to predict Pask's ratings, the R-squared value was much lower, something like 0.2. I also did this with the 'match score' ratings. So by themselves they would not provide a good model, but when combined it does increase the adjusted R-squared by nearly 0.1.

"The output of the linear regression module for the full model is the scaled ACF difficulty, and these coefficients are used directly to generate the 11-man difficulty ratings."

null

So there you have it. 11-man ballot has at long last been put on a solid footing based on detailed calculation and computer analysis. Is this the very last word on the topic? Ed himself would likely say that in checkers, the last word is never really in, but have we ever come a long way!

And some people think checkers is simple? After hundreds of years, we're still finding new ideas and exciting discoveries!


The Checker Maven thanks Ed Gilbert for the enormous privilege of presenting his work in this column. We hope Mr. Gilbert's ground-breaking efforts will serve to increase interest in 11-man ballot checkers, leading to more competition and even more analysis and publication.

Main link to Ed's work:

11-man ballot, descriptive page

The following links can also be reached from Ed's website but are given here for convenience.

Latest KingsRow which includes the 11-man ballot opening book

PDN for all 11-man ballots

PDN for 11-man drawn ballots

PDN for 11-man lost ballots

11-man ballot difficulty ratings in order of difficulty

11-man ballot difficult ratings in ballot number order

Reference rankings of 3-move opening difficulty (Richard Pask/ACF)

null



Solution

You can play over the solution on your board or computer, or view a full animation here.

Play continues after the balloted moves, from the diagram above.


1. 16-20 25-22
2. 1-6 29-25
3. 11-15 18x11
4. 8x15 23-18
5. 7-11 26-23
6. 9-13 24-19
7. 15x24 28x19
8. 2-7 31-26
9. 4-8 32-28
10. 8-12 28-24
11. 3-8 18-15
12. 11x18 22x15
13. 5-9 26-22
14. 9-14 22-18
15. 14-17 25-21
16. 17-22 18-14
17. 10x17 21x14
18. 6-9! 14x5
19. 22-26 30-25
20. 7-11 ...

The last of Black's 20 (count 'em) star moves.


20. ... 15-10

Black now has a man-down draw. Play might continue as follows.


21. 26-30 25-22
22. 11-15 23-18
23. 8-11 10-7
24. 13-17! 22x13
25. 15x22 5-1
26. 11-16 19-15
27. 16-19 7-3
28. 19x28 ...

Drawn. Have you ever seen such a thin draw?

null

09/14/19 - Category: General -Printer friendly version-
You can email the Webmaster with comments on this article.