Return to my Checkers pages
Go to my home page
© Copyright 1997, Jim Loy
You may print this and show it to others. But, this article will eventually be
part of a book that I am writing. So, please do not distribute it widely.
If you need help reading checkers notation, please print out the numbered board.
This position is
at the end of the game Bryden-Kirkind. Red is a piece ahead, and moves 3-7 to
stay a piece ahead. White resigns, here. Continue 6-1 (6-2 10-15 RW) 7-11 RW.
The program that I use, Sage 4001, completely botches this position, at less
than 5 minutes per move: 24-27? 6-1 27-31 1-6 12-16 6-15 16-19 draw.
At the diagram, Sage sees that Red will have two kings and one piece, while White will have three pieces. A study of 5 minutes shows Sage that White will king one piece (a human can see that all three will king). And so, Sage will choose the correct move at a skill-level of 5 minutes or more.
Sage thinks that kings are very valuable. And so they are, sometimes. Here, a superficial glance at the board shows a human that there is no way to stop White from kinging several pieces. The computer should not be fooled this easily.
Perhaps kings are valued too heavily, in this position. Perhaps potential kings are valued too lightly. Perhaps the search should be extended, when there are potential kings which may be unstoppable.
This position
will probably result from Ballantyne-Buchanan. It is
a simple draw. But, Sage cannot draw it at any skill-level. It should be easy
for the computer.
This position is fairly funny, when Sage is asked to move. 13-9?? 19-23 (Sage moves 19-24??) 9-14 23-27?? (23-26 is the win). Sage moves the kings around forever, never preventing the winning move, and never making the winning move.
What is the problem here? A problem is that Sage does not treat a repeated position (during look-ahead) as a draw. During the game, a repeated position is not a draw. But, during look-ahead, it should be an absolute draw. That may not be the problem here, as Sage does not seem to know that (after the 23-26 pitch) this 3 piece vs. 2 piece ending is a win for Red. And it does have Chinook's 5-piece ending database. It would seem that Sage is not using the database in this position.
There are lots of
positions that Sage evaluates incorrectly, because of repeated positions. Here
is one (which doesn't lead to any disaster) from Kear-Richmond. Sage thinks
that Red has a fairly strong advantage, and yet it repeats the position forever
(as it should). Sage should be able to tell that this position is a draw.
So what do I care if Sage evaluates the position incorrectly? It still makes the right moves here. But, if Sage had known that this was a draw, then maybe it could have avoided the position, and won. This happens in similar positions.
At the very least, evaluating repeated positions as draws saves time. Sage looks at the same position over and over during a single search. Sage should know that it has seen this position earlier, in the same search, and know that it is not making any progress.
Many positions are just too complicated to evaluate correctly at a given skill-level. And we can't blame Sage when it does that. Many man-down endings, the difficult mail-play openings, and gambits like the White Doctor are routinely mis-evaluated by the programs (and by humans too). That's not a problem.
Note: I would like to hear about positions that other programs mis-evaluate, especially positions which are simple for a human to evaluate.
Addendum:
I just spent a lot of time on an experiment with Nexus, the program that I am currently using. I set up First Position, with Red's King in the 32, 28 double corner and the Red piece on 3. Nexus knows that this is a WW. I have set Nexus on "infinite mode." Well, a half-hour and over 1000 moves later, Nexus has still not forced Red's king out of the double corner or forced the piece to move off 3. This is a flaw in the Chinook database. The computer does not ever allow a drawn position. But it has no way of knowing (without looking ahead many moves), which positions are closer to winning a piece. First Position and similar "difficult" positions are the main reason for having a huge database in the first place. In this very common ending, the database fails completely.