Author: Todd Durham
Date: 19:52:55 03/21/99
Go up one level in this thread
On March 21, 1999 at 21:04:55, James T. Walker wrote:
>I played 100 games with Crafty 16.5 vs Crafty 16.5 with all 3/4 man tablebases
>and a few 5 man tablebases in an attempt to determine how much help the
>tablebases are. Has anyone else done this and if so, what was the result?
>The final score was Crafty w/tb +32 =38 -30 or 51 to 49. The first 50 games
>were very different from the second 50 games even though the match was
>uninterrupted for the 100 games.
> First 50 Crafty w/tb +21 =15 -14
> Second 50 " " +11 =23 -16
>It looks like two different programs. Just shows that even 50 games is not a
>large enough sample to prove anything. I still don't know what the ELO gain is
>from tablebases ! I'm planning on a couple hundred more games later this week.
>If anyone else has done this (Bob?) please post your results.
>Jim Walker
>P.S. These were only Game/5 minutes.
Hyatt wrote an interesting post on rgcc just yesterday that relates to this. I
give it in full, with apologies to all if it is bad etiquette to post someone
else's stuff. (I'm new here, and if this is bad etiquette, someone please tell
me.)
Author:
Robert Hyatt <hyatt@crafty.cis.uab.edu>
Date:
1999/03/20
Forum:
rec.games.chess.computer
Gene Ward Smith <gsmith@blazenetme.net> wrote:
: In article <7cj7fl$g20$2@juniper.cis.uab.edu>, Robert Hyatt
: <hyatt@crafty.cis.uab.edu> wrote:
:>Akorps <akorps@aol.com> wrote:
:>: Is there going to be a web page set up
:>: where one can play out positions from
:>: these tablebases? There used to be one
:>: but it seems to have been taken down
:>: (was extremely instructive)
:>This is doable, but I can't do it. Our ftp machine typically has 50-60
:>people downloading 24 hours per day. Not any real way to also allow
:>multiple copies of Crafty to run and play games, even if they start off
:>in the tablebases.
: It would be very interesting to know, given these tablebases, how many
: endgames can be played correctly by means of heuristics. By
: "correctly" I don't mean always obtaining the fastest possible mate,
: but at least winning in won positions, and not losing in drawn ones.
: --
: Gene Ward Smith
: gsmith@blazenetme.net
I can give you one that was terribly surprising. we had a long discussion
about this on CCC, and Mark Young (I think) reported that Fritz could win
KQ vs KR at some fairly quick time control (maybe 30 seconds per move or
something, I am not certain). I was definitely skeptical, because this is
a tricky ending for a human. So I tried a bunch of random KQ vs KR positions,
using crafty for both sides. The white side had the KQ and _no_ tablebases,
the black side had KR + all tablebases. And I found that even at 1 second
per move, crafty could win that, meaning that the simple heuristic of driving
the losing king to the edge and then to the corner was enough to let a
search find the win. I was surprised, because the tablebases have some wild
defensive moves that I thought would make it fail. Notice that white did not
win _perfectly_, but it _always_ won within the allowed 50 moves. I have also
run a similar test with KBN vs K, although not recently, and again, the
simple heuristic of driving the king to the edge, then the _right_ corner is
enough that a fairly short search can force the mate. Not perfectly, but
within 50 moves, which is all that counts.
Other endings are harder, ie KRP vs KR. I have special eval code for this,
but it is not 'perfect' and the databases will find clever ways to 'break'
this test.
--
Robert Hyatt Computer and Information Sciences
hyatt@cis.uab.edu University of Alabama at Birmingham
(205) 934-2213 115A Campbell Hall, UAB Station
(205) 934-5473 FAX Birmingham, AL 35294-1170
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.