Bratko Kopec #2
brucemo
This is position #2 from the Bratko-Kopec test suite:
3r1k2/4npp1/1ppr3p/p6P/P2PPPP1/1NR5/5K2/2R5 w - - 0 1
a b c d e f g h
+---+---+---+---+---+---+---+---+
8 | |///| |/r/| |/k/| |///| 8
+---+---+---+---+---+---+---+---+
7 |///| |///| |/n/| p |/p/| | 7
+---+---+---+---+---+---+---+---+
6 | |/p/| p |/r/| |///| |/p/| 6
+---+---+---+---+---+---+---+---+
5 |/p/| |///| |///| |///| P | 5
+---+---+---+---+---+---+---+---+
4 | P |///| |/P/| P |/P/| P |///| 4
+---+---+---+---+---+---+---+---+
3 |///| N |/R/| |///| |///| | 3
+---+---+---+---+---+---+---+---+
2 | |///| |///| |/K/| |///| 2
+---+---+---+---+---+---+---+---+
1 |///| |/R/| |///| |///| | 1
+---+---+---+---+---+---+---+---+
a b c d e f g h
The problem is white to play and win. The position is from a game
between Evfim Bogolyubov and Rudolf Spielmann, 1932. It got into
the test suite because Hans Kmoch discussed this in "Pawn Power in
Chess". In my edition the diagram appears on page 183, in the "The
Sealer and the Sweeper" chapter.
The position is interesting to chess players because the key move
is a positional pawn sacrifice. It is interesting to computer
folks because it is obvious after the key move that white gets
something for his pawn, but programs have difficulty understanding
what. Programs like to play Kf3 or f5 or e5.
The key move is 1. d5, and the main line, as given by Kmoch, goes
as follows:
1. d5! cxd5
2. e5! R6d7
3. Nd4
In the resulting position, white is a pawn down, but has secured
approximately five advantages:
1) Black has a backward pawn on the b-file.
2) White has control of the open c-file.
3) Black's pressure on the half-open d-file has been negated.
4) White has a nice knight on d4.
5) White has achieved an advantageous pawn formation known as a
"quart-grip" on the e..h files.
The game was eventually a draw, but Kmoch makes the reasonable
claim that a win was easy from this position.
Up until now I have never heard of anyone who claimed that their
program solved this position. There is a chart on p. 223 of
"Computers, Chess, and Cognition", ed. by Marsland and Shaeffer,
1990, that shows that all of twenty-one programs that attempted
this position failed to find the key. And the programs that failed
at this aren't wimps either, included in those who failed are Deep
Thought, Cray Blitz, Hitech, and some top micros of the day,
Mephisto and Rebel, plus other programs that did well in events,
including some that had a lot of knowledge.
I did an experiment to see if the problem here is search depth,
since the test has a time limit of two minutes per position. I let
my program, which is no pushover, especially in the ending, have a
go at this for about nine hours (it is still going). It is
currently stuck in the 19th iteration, and thinks 1. f5 is +0.50.
Today I purchased a copy of the 11-Nov-96 issue of "Inside Chess".
On page 23 of this magazine is an ICD ad for Rebel 8.0, which
includes a claim (by CCR editor Enrique Irazoqui) that Rebel 8.0 is
the only "top program" that can solve this, and that this "proves
its ability to sacrifice material for positional advantage".
I do not dispute these claims. I would like to ask Ed if Rebel
finds a win of material after 1. d5, or did it solve this problem
on "positional" grounds? If it solves it positionally, what
elements of white's advantage is Rebel capable of seeing?
My own program has understanding of approximately three and a half
of white's positional advantages, but it also sees a false
positional advantage for black -- black's brand-new d-pawn is
passed, is on a central square, and has two rooks behind it. That
the pawn is contained is something it doesn't understand. The sum
total of all of this ends up being about -0.45 after ten minutes or
so, if I let it evaluate after 2. ... R6d7.
I will understand if Ed doesn't want to divulge trade secrets by
answering this question, and I apologize for putting him on the
spot by asking. I ask only because I believe that solving this
problem is something monumental, as it is a very old and very
difficult problem.
I would be delighted to hear any other discussion of this position,
and of other programmers' successful attempts to deal with it :-)
bruce
Robert Hyatt
brucemo (bru...@nwlink.com) wrote:
: This is position #2 from the Bratko-Kopec test suite:
Bruce and I played with this last nite... Crafty thinks the best move is
+.1 for white, and if told to search d5, thinks that is -.1, which is
interesting to me because it too sees that this is a pawn sacrifice.
I'm not yet sure why it likes the position, although it does understand
that a blockaded pawn with rooks behind it is no big deal. More after
I analyze this in depth...
Bob
Howard Exner
This position reminds me of another position which Rebel 8.0
finds while others do not. Similar in that Rebel 8.0 sees that
its Rooks will own the d file by allowing itself the positional
"no no" of doubled c pawns. (In the Bratko example it gives up a
pawn for rook ownership of the c file as well as the other pluses
pointed out be Bruce Moreland). I wonder if the active rooks is
the knowledge component that boosts the eval most?
dxc5 although resulting in doubled c pawns will enable white
to activate his rooks on d5 and d1.
r3r1k1/pp3pp1/2np1q1p/2p1p3/2PPQ3/2P1PB2/P4PPP/R3R1K1 w;Botvinnik-
Chekhover(Leningrad,1938)
How do Crafty, Ferret and other programs play this
positional example?
Martin Borriss
In article <328A1A...@nwlink.com>,
As usual, Bruce's post was very interesting.
When I first saw this position, I was impressed by the solution move d5. I
would definitely consider it in a practical game, particulary if a win is
needed.
OTOH, to me it is not clear whether missing the solution is really a fault
in a program. Other moves keep the small advantage white has in this
position, but probably won't win. My problem is that I cannot see the _win_
after the pawn sacrifice. Maybe missing d5 is not a problem at all ?!
Perhaps it would be an idea to let Ferret and Crafty play each other in the
position after the pawn sac ?
Martin
Ernst A. Heinz
> I would be delighted to hear any other discussion of this position,
> and of other programmers' successful attempts to deal with it :-)
>
> bruce
Hi Bruce,
DarkThought/Jakarta solves BK #2 easily. On a 275MHz Alpha-21064a (approx.
1.5x slower than a 200MHz Pentium-Pro) it gives us the following:
+---+---+---+---+---+---+---+---+
8 | |:::| |*R:| |*K:| |:::|
+---+---+---+---+---+---+---+---+
7 |:::| |:::| |*N:|*P |*P:| |
+---+---+---+---+---+---+---+---+
6 | |*P:|*P |*R:| |:::| |*P:|
+---+---+---+---+---+---+---+---+
5 |*P:| |:::| |:::| |:::| P |
+---+---+---+---+---+---+---+---+
4 | P |:::| |:P:| P |:P:| P |:::|
+---+---+---+---+---+---+---+---+
3 |:::| N |:R:| |:::| |:::| |
+---+---+---+---+---+---+---+---+
2 | |:::| |:::| |:K:| |:::|
+---+---+---+---+---+---+---+---+
1 |:::| |:R:| |:::| |:::| |
+---+---+---+---+---+---+---+---+
A B C D E F G H
White moves d5
Engine plays White.
# Starting to search.
* 33 legal moves.
* Root eval = 26 (0.20), material balance = 0 (0.00).
* Fixed search time = 00:01:00.
* Timer interrupt in 60 sec.
00:00:00 6.01[ -17, 63] c3c2 d6d7 f2f3 f8e8 c2c3 d7d6 =27 (0.21) #66031
00:00:01 6.13[ 27, 28] d4d5 d6f6 f4f5 g7g6 h5g6 f6e6 >=28! (0.22) #81823
00:00:01 7.01[ -4, 76] d4d5 c6d5 e4e5 d6d7 b3d4 e7g8 e5e6 g8f6 =33 (0.26) #124607
00:00:02 8.01[ -7, 73] d4d5 c6d5 e4e5 d6d7 b3d4 d8e8 e5e6 f7e6 d4e6 f8f7 =37 (0.29) #207973
00:00:05 9.01[ 5, 85] d4d5 c6d5 e4e5 d6d7 b3d4 d8e8 c1b1 e8c8 e5e6 f7e6 c3c8 e7c8
d4e6 f8f7 =52 (0.41) #516208
00:00:12 10.01[ 12, 92] d4d5 c6d5 e4e5 d6d7 b3d4 d7b7 e5e6 d8e8 e6f7 f8f7 c3c2 b7b8
=50 (0.39) #1058613
00:00:27 11.01[ 18, 98] d4d5 c6d5 e4e5 d6d7 b3d4 d7b7 e5e6 d8e8 e6f7 f8f7 c1e1 e7g8
e1b1 =56 (0.44) #2751425
=Ernst=
+----------------------------------------------------------------------------+
| Ernst A. Heinz, School of CS (IPD), Univ. of Karlsruhe, P.O. Box 6980, |
| D-76128 Karlsruhe, F.R. Germany. WWW: <http://wwwipd.ira.uka.de/~heinze> |
| Mail: <hei...@ira.uka.de> Tel: +49-(0)721-6084386 Fax: +49-(0)721-694092 |
+----------------------------------------------------------------------------+
"It has recently been found out that research causes cancer in rats!"
Ed Schröder
From: brucemo <bru...@nwlink.com>
: This is position #2 from the Bratko-Kopec test suite:
: 3r1k2/4npp1/1ppr3p/p6P/P2PPPP1/1NR5/5K2/2R5 w - - 0 1
[ snip ]
: I do not dispute these claims. I would like to ask Ed if Rebel
: finds a win of material after 1. d5, or did it solve this problem
: on "positional" grounds? If it solves it positionally, what
: elements of white's advantage is Rebel capable of seeing?
: bruce
After 6 plys Rebel plays 1.d5
After 12 plys still 1.d5
No gain of material is seen...
Score is about +0.70
After 1.d5 cxd5 2.e5 R6d7 3.Nd4 the following factors are important
for Rebel:
1.. complete control of the c-file
2.. the weak black pawns on b6 and d5
3.. the white knight on d4
4.. the position of the white king in the center.
5.. the superior white pawn formation on the king side.
This is more than sufficient for a pawn sacrifice.
However I know dozen of such positions where Rebel does not find
the correct pawn sacrifice. The NOLUT positions comes in mind.
- Ed Schroder -
BTW I checked Rebel Decade too and RD plays 1.d5 even from the 3th ply.
brucemo
> >
>
>
> When I first saw this position, I was impressed by the solution move d5. I
> would definitely consider it in a practical game, particulary if a win is
> needed.
>
> OTOH, to me it is not clear whether missing the solution is really a fault
> in a program. Other moves keep the small advantage white has in this
> position, but probably won't win. My problem is that I cannot see the _win_
> after the pawn sacrifice. Maybe missing d5 is not a problem at all ?!
>
> Perhaps it would be an idea to let Ferret and Crafty play each other in the
> position after the pawn sac ?
The winning line, given by Kmoch, is:
1. d5! cxd5
2. e5! R6d7
3. Nd4 Rb8
4. f5 R7d8
5. R1c2 Re8
6. Ke3 Rbc8
7. Rxc8 Nxc8
8. Kd3
"with a rather easy win".
In the game white erred with:
7. Nb5?? Rxc2+
8. Rxc2 Nxf5+!
9. gxf5 Rxe5+
10.Kf4 Re4+
11.Kg3 Rxa4
12.f6 Rc4!
13.Re3 Re4!
14.Rc3 Rc4 =
I have two computers working on this now. One is analyzing the root position,
it has completed 19 plies, and has spent the last 24 hours at least trying to
resolve the PV move at ply 20. It hasn't found 1. d5 yet, but has switched to
1. e5, after which the state of d5 makes me wince. Its line is: e5 R6d7 Kf3
Rc8 f5 Rdc7 Nd2 Nd5 Rc4 Ke7 Ne4 Nb4 Nd6 Rd8 Ke3 Rh8 Ke4 Rd7 R1c3 +0.53.
The other one is looking at the position after 2. ... R6d7, and has spent the
last 24 hours trying to find a PV move at ply 19. It finds 3. Nd4 right away
of course, so what I'm looking for is the eval it comes up with. This
computer believes it has found a repetition draw. Its line is: Nd4 Rb8 Rc7
Ke8 e6 Rxc7 Rxc7 Rc8 Rb7 Rc4 Nb5 Rc2+ Ke3 fxe6 Nd6+ Kd8 Nf7+ Ke8 Nd6+ -0.01
(draw).
It might be interesting to arrange a match, yes, but I think your comments are
more useful :-)
bruce
brucemo
Ed Schröder wrote:
>
> From: brucemo <bru...@nwlink.com>
>
> : This is position #2 from the Bratko-Kopec test suite:
>
> : 3r1k2/4npp1/1ppr3p/p6P/P2PPPP1/1NR5/5K2/2R5 w - - 0 1
> After 6 plys Rebel plays 1.d5
> After 12 plys still 1.d5
> No gain of material is seen...
> Score is about +0.70
>
> After 1.d5 cxd5 2.e5 R6d7 3.Nd4 the following factors are important
> for Rebel:
>
> 1.. complete control of the c-file
> 2.. the weak black pawns on b6 and d5
> 3.. the white knight on d4
> 4.. the position of the white king in the center.
> 5.. the superior white pawn formation on the king side.
>
> This is more than sufficient for a pawn sacrifice.
>
> However I know dozen of such positions where Rebel does not find
> the correct pawn sacrifice. The NOLUT positions comes in mind.
You mean that small test suite full of violently hard problems?
Thanks for the explaining Rebel's thinking about this problem. Good
job solving it :-)
bruce
brucemo
Howard Exner wrote:
>[snip]
> dxc5 although resulting in doubled c pawns will enable white
> to activate his rooks on d5 and d1.
>
> r3r1k1/pp3pp1/2np1q1p/2p1p3/2PPQ3/2P1PB2/P4PPP/R3R1K1 w;Botvinnik-
> Chekhover(Leningrad,1938)
>
> How do Crafty, Ferret and other programs play this
> positional example?
Cool position. After a minute or two, and probably permanently, mine doesn't
like to play dxc5, but it's just a matter of a few centipawns difference.
bruce
Martin Borriss
In article <56et79$r...@nz12.rz.uni-karlsruhe.de>,
hei...@ira.uka.de (Ernst A. Heinz) writes:
>
>DarkThought/Jakarta solves BK #2 easily. On a 275MHz Alpha-21064a (approx.
>1.5x slower than a 200MHz Pentium-Pro) it gives us the following:
>
> +---+---+---+---+---+---+---+---+
>8 | |:::| |*R:| |*K:| |:::|
> +---+---+---+---+---+---+---+---+
>7 |:::| |:::| |*N:|*P |*P:| |
> +---+---+---+---+---+---+---+---+
>6 | |*P:|*P |*R:| |:::| |*P:|
> +---+---+---+---+---+---+---+---+
>5 |*P:| |:::| |:::| |:::| P |
> +---+---+---+---+---+---+---+---+
>4 | P |:::| |:P:| P |:P:| P |:::|
> +---+---+---+---+---+---+---+---+
>3 |:::| N |:R:| |:::| |:::| |
> +---+---+---+---+---+---+---+---+
>2 | |:::| |:::| |:K:| |:::|
> +---+---+---+---+---+---+---+---+
>1 |:::| |:R:| |:::| |:::| |
> +---+---+---+---+---+---+---+---+
> A B C D E F G H
>00:00:00 6.01[ -17, 63] c3c2 d6d7 f2f3 f8e8 c2c3 d7d6 =27 (0.21) #66031
>00:00:01 6.13[ 27, 28] d4d5 d6f6 f4f5 g7g6 h5g6 f6e6 >=28! (0.22) #81823
>00:00:01 7.01[ -4, 76] d4d5 c6d5 e4e5 d6d7 b3d4 e7g8 e5e6 g8f6 =33 (0.26) #124607
>00:00:02 8.01[ -7, 73] d4d5 c6d5 e4e5 d6d7 b3d4 d8e8 e5e6 f7e6 d4e6 f8f7 =37 (0.29) #207973
>00:00:05 9.01[ 5, 85] d4d5 c6d5 e4e5 d6d7 b3d4 d8e8 c1b1 e8c8 e5e6 f7e6 c3c8 e7c8
> d4e6 f8f7 =52 (0.41) #516208
>00:00:12 10.01[ 12, 92] d4d5 c6d5 e4e5 d6d7 b3d4 d7b7 e5e6 d8e8 e6f7 f8f7 c3c2 b7b8
> =50 (0.39) #1058613
>00:00:27 11.01[ 18, 98] d4d5 c6d5 e4e5 d6d7 b3d4 d7b7 e5e6 d8e8 e6f7 f8f7 c1e1 e7g8
> e1b1 =56 (0.44) #2751425
It plays d5 because it thinks it has 1.44 pawns positional advantage in the
final position of the last PV. This seems way too much. e5e6 and e6f7 are not
best here and probably cannot justify d5...
Martin
Martin Borriss
In article <328B5F...@nwlink.com>,
brucemo <bru...@nwlink.com> writes:
>Martin Borriss wrote:
>>
>> In article <328A1A...@nwlink.com>,
>> brucemo <bru...@nwlink.com> writes:
>> >
>> in a program. Other moves keep the small advantage white has in this
>> position, but probably won't win. My problem is that I cannot see the _win_
>> after the pawn sacrifice. Maybe missing d5 is not a problem at all ?!
>>
>
>
>1. d5! cxd5
>2. e5! R6d7
>3. Nd4 Rb8
>4. f5 R7d8
>5. R1c2 Re8
>6. Ke3 Rbc8
>7. Rxc8 Nxc8
>8. Kd3
>
>"with a rather easy win".
>
I see. 5... Re8 is doubtful, although it paid off in the game.
I think that many of those book lines have to be checked very carefully.
E.g. Ferret will probably defend such a position as good as say Kmoch.
That's why I wanted to see a program win with this sacrifice against another
program first.
A good plan might be 3.Nd4, 4.f5, 5.Ke3, 6. Kf4 with the idea e6 followed
by Ke5 (to provoke f7-f6). White is very lucky here, everything fits
together.
Programs which have adjusted their eval based on this position run a huge
risk to give pawns in similar positions where they really are not getting
much in return. That is why I am satisfied with Gullydeckel playing 1.Ke3
here :)
Martin
>In the game white erred with:
>
>7. Nb5?? Rxc2+
>8. Rxc2 Nxf5+!
>9. gxf5 Rxe5+
>10.Kf4 Re4+
>11.Kg3 Rxa4
>12.f6 Rc4!
>13.Re3 Re4!
>14.Rc3 Rc4 =
>
>I have two computers working on this now. One is analyzing the root position,
>it has completed 19 plies, and has spent the last 24 hours at least trying to
>resolve the PV move at ply 20. It hasn't found 1. d5 yet, but has switched to
>1. e5, after which the state of d5 makes me wince. Its line is: e5 R6d7 Kf3
>Rc8 f5 Rdc7 Nd2 Nd5 Rc4 Ke7 Ne4 Nb4 Nd6 Rd8 Ke3 Rh8 Ke4 Rd7 R1c3 +0.53.
>
>The other one is looking at the position after 2. ... R6d7, and has spent the
>last 24 hours trying to find a PV move at ply 19. It finds 3. Nd4 right away
>of course, so what I'm looking for is the eval it comes up with. This
>computer believes it has found a repetition draw. Its line is: Nd4 Rb8 Rc7
>Ke8 e6 Rxc7 Rxc7 Rc8 Rb7 Rc4 Nb5 Rc2+ Ke3 fxe6 Nd6+ Kd8 Nf7+ Ke8 Nd6+ -0.01
>(draw).
>
en...@bitmailer.net
In article ,
bor...@os.inf.tu-dresden.de (Martin Borriss) wrote:
>
> In article ,
> brucemo writes:
> >Up until now I have never heard of anyone who claimed that their
> >program solved this position. There is a chart on p. 223 of
> >"Computers, Chess, and Cognition", ed. by Marsland and Shaeffer,
> >1990, that shows that all of twenty-one programs that attempted
> >this position failed to find the key. And the programs that failed
> >at this aren't wimps either, included in those who failed are Deep
> >Thought, Cray Blitz, Hitech, and some top micros of the day,
> >Mephisto and Rebel, plus other programs that did well in events,
> >including some that had a lot of knowledge.
> >
Rebel 8 is the only program I know that solves this position, and almost instantenously!
Enrique
-----------------------------------------------------------------------
This article was posted to Usenet via the Posting Service at Deja News:
http://www.dejanews.com/ [Search, Post, and Read Usenet News]
john quill taylor
brucemo <bru...@nwlink.com> wrote:
>This is position #2 from the Bratko-Kopec test suite:
>3r1k2/4npp1/1ppr3p/p6P/P2PPPP1/1NR5/5K2/2R5 w - - 0 1
> a b c d e f g h
> +---+---+---+---+---+---+---+---+
>8 | |///| |/r/| |/k/| |///| 8
> +---+---+---+---+---+---+---+---+
>7 |///| |///| |/n/| p |/p/| | 7
> +---+---+---+---+---+---+---+---+
>6 | |/p/| p |/r/| |///| |/p/| 6
> +---+---+---+---+---+---+---+---+
>5 |/p/| |///| |///| |///| P | 5
> +---+---+---+---+---+---+---+---+
>4 | P |///| |/P/| P |/P/| P |///| 4
> +---+---+---+---+---+---+---+---+
>3 |///| N |/R/| |///| |///| | 3
> +---+---+---+---+---+---+---+---+
>2 | |///| |///| |/K/| |///| 2
> +---+---+---+---+---+---+---+---+
>1 |///| |/R/| |///| |///| | 1
> +---+---+---+---+---+---+---+---+
> a b c d e f g h
>The problem is white to play and win.
>...
I ran MChess Pro 5.0 on this through completion of 13 ply, for 15 hours,
and after 1.6 billion nodes searched, with a value of +0.18 it likes:
1.f5 R6d7 2.Kg3 Rc8 3.d5 Rcd8 4.dxc6 Nxc6 5.Rxc6 Rd3+ 6.Kf4 and so on.
It found 1.f5 immediately, then switched to 1.Kf3 at 56 seconds, but
then it switched back to 1.f5 permanently at 39 minutes.
__
john quill taylor / /\
writer at large / / \
Hewlett-Packard, Storage Systems Division __ /_/ /\ \
Boise, Idaho U.S.A. /_/\ __\ \ \_\ \
e-mail: jqta...@hpdmd48.boi.hp.com \ \ \/ /\\ \ \/ /
Telephone: (208) 396-2328 (MST = GMT - 7) \ \ \/ \\ \ /
Snail Mail: Hewlett-Packard \ \ /\ \\ \ \
11413 Chinden Blvd \ \ \ \ \\ \ \
Boise, Idaho 83714 \ \ \_\/ \ \ \
Mailstop 852 \ \ \ \_\/
\_\/
"When in doubt, do as doubters do." - jqt -
haiti, rwanda, cuba, bosnia, ... we have a list,
where is our schindler?
brucemo
Martin Borriss wrote:
> Programs which have adjusted their eval based on this position run a huge
> risk to give pawns in similar positions where they really are not getting
> much in return. That is why I am satisfied with Gullydeckel playing 1.Ke3
> here :)
I think this is a fine attitude, although I think that it is also a good
thing to try to have material and positional terms balanced in such a way
that a program does attempt such enterprises as 1. d5 occasionally :-)
bruce
brucemo
Martin Borriss wrote:
> I see. 5... Re8 is doubtful, although it paid off in the game.
> I think that many of those book lines have to be checked very carefully.
> E.g. Ferret will probably defend such a position as good as say Kmoch.
> That's why I wanted to see a program win with this sacrifice against another
> program first.
Here are two games for you. The first is a playoff between my program and itself, at a time control
of 30 0, on a P6/200, after the moves 1. d5 cxd5 2. e5 R6d7 have been played. The second is the
same deal only I didn't fake any moves in first (I was very surprised at 1. Rb1, normally it play
s 1. Ke3 or 1. f5). My apologies for PGN adherents, as I don't know how to specify a FEN string in
a PGN file, I just guessed.
It was easier to do this than to do Ferret vs Crafty. As time allows I think I'd be willing to play
either side of this vs anyone, with or without those initial four moves as prelude. No special
opening books please! :-)
I draw no conclusions from either of these games, and make no excuses for any poor play you may
expose yourself to if you play over these :-)
If you play over the second game, you should know that a KN vs KP database was active at the end.
bruce
----
[Event "Bratko-Kopec #2 playoff"]
[Site "Basement"]
[Date "15-Nov-96"]
[Round "1"]
[White "Ferret"]
[Black "Ferret"]
[Result "1/2-1/2"]
[Fen "3r1k2/3rnpp1/1p5p/p2pP2P/P4PP1/1NR5/5K2/2R5 w - - 0 1"]
1. Nd4 Rb8 2. Nb5 Rbd8 3. Nd6 f6 4. Rb3 d4 5. Rxb6 Nd5 6. Rb7 Rxb7 7. Nxb7 Rb8
8. Nc5 Ke7 9. Kf3 Nc3 10. Re1 Rd8 11. Nd3 Rd5 12. exf6+ Kxf6 13. f5 Nxa4 14.
Re6+ Kf7 15. Ne5+ Kf8 16. Ng6+ Kf7 17. Ne5+ Kf8 18. Ng6+ Kf7 19. Ne5+ 1/2-1/2
[Event "Bratko-Kopec #2 playoff"]
[Site "Basement"]
[Date "15-Nov-96"]
[Round "2"]
[White "Ferret"]
[Black "Ferret"]
[Result "1/2-1/2"]
[Fen "3r1k2/4npp1/1ppr3p/p6P/P2PPPP1/1NR5/5K2/2R5 w - - 0 1"]
1. Rb1 Kg8 2. Ke3 Re8 3. Kf3 Kf8 4. g5 Rd6 5. Kg4 Rc8 6. Rb2 Ke8 7. gxh6 f5+
8. exf5 gxh6 9. Re3 Kf7 10. Re5 Nd5 11. Kf3 Nf6 12. Rbe2 Rc7 13. Re6 Rcd7 14.
Rc2 Rxe6 15. fxe6+ Kxe6 16. Rxc6+ Rd6 17. f5+ Kd5 18. Rxd6+ Kxd6 19. Nd2 Kd5
20. Ke3 Nxh5 21. Ne4 Ng7 22. Ng3 Ne8 23. Nh5 Nd6 24. Nf4+ Kc4 25. f6 b5 26.
axb5 Kxb5 27. Kd3 Kc6 28. Kc2 Kd7 29. Kb3 Ke8 30. Ka4 Kf7 31. Kxa5 Kxf6 32.
Kb4 Nf5 33. Ne2 Nxd4 34. Nxd4 Kg5 35. Kc4 h5 36. Kb4 Kf4 37. Kc4 h4 38. Ne2+
Ke3 39. Ng1 Kf2 40. Nh3+ Kg3 41. Ng1 Kf2 42. Nh3+ Kg3 43. Ng1 Kf2 1/2-1/2
brucemo
Ernst A. Heinz wrote:
>
> > I would be delighted to hear any other discussion of this position,
> > and of other programmers' successful attempts to deal with it :-)
> >
> > bruce
>
> Hi Bruce,
>
> 1.5x slower than a 200MHz Pentium-Pro) it gives us the following:
>
> +---+---+---+---+---+---+---+---+
> 8 | |:::| |*R:| |*K:| |:::|
> +---+---+---+---+---+---+---+---+
> 7 |:::| |:::| |*N:|*P |*P:| |
> +---+---+---+---+---+---+---+---+
> 6 | |*P:|*P |*R:| |:::| |*P:|
> +---+---+---+---+---+---+---+---+
> 5 |*P:| |:::| |:::| |:::| P |
> +---+---+---+---+---+---+---+---+
> 4 | P |:::| |:P:| P |:P:| P |:::|
> +---+---+---+---+---+---+---+---+
> 3 |:::| N |:R:| |:::| |:::| |
> +---+---+---+---+---+---+---+---+
> 2 | |:::| |:::| |:K:| |:::|
> +---+---+---+---+---+---+---+---+
> 1 |:::| |:R:| |:::| |:::| |
> +---+---+---+---+---+---+---+---+
> A B C D E F G H
> Engine plays White.
> # Starting to search.
> * 33 legal moves.
> * Root eval = 26 (0.20), material balance = 0 (0.00).
> * Fixed search time = 00:01:00.
> * Timer interrupt in 60 sec.
> 00:00:01 6.13[ 27, 28] d4d5 d6f6 f4f5 g7g6 h5g6 f6e6 >=28! (0.22) #81823
> 00:00:01 7.01[ -4, 76] d4d5 c6d5 e4e5 d6d7 b3d4 e7g8 e5e6 g8f6 =33 (0.26) #124607
> 00:00:02 8.01[ -7, 73] d4d5 c6d5 e4e5 d6d7 b3d4 d8e8 e5e6 f7e6 d4e6 f8f7 =37 (0.29) #207973
> 00:00:05 9.01[ 5, 85] d4d5 c6d5 e4e5 d6d7 b3d4 d8e8 c1b1 e8c8 e5e6 f7e6 c3c8 e7c8
> d4e6 f8f7 =52 (0.41) #516208
> 00:00:12 10.01[ 12, 92] d4d5 c6d5 e4e5 d6d7 b3d4 d7b7 e5e6 d8e8 e6f7 f8f7 c3c2 b7b8
> =50 (0.39) #1058613
> 00:00:27 11.01[ 18, 98] d4d5 c6d5 e4e5 d6d7 b3d4 d7b7 e5e6 d8e8 e6f7 f8f7 c1e1 e7g8
> e1b1 =56 (0.44) #2751425
What interests me is not the ply 11 PV, but the ply 6 and 7 PV's. It's fascinating
that there'd be enough compensation so early. I wonder if the compensation is evident
if you do an eval of the position after 1. d5 cxd5 2. e5 R6d7 3. Nd4 ?
Please don't take this the wrong way, but did you guys ever tune for this position or
did this just happen? I know this sounds like an accusation, but I don't mean it to
be. I would have no problems tuning for this position since there are a lot of
positional aspects here that do deserve a big bonus. I made a half-hearted attempt to
tune for this, by trying to analyze some aspects of the position that deserved
attention in the form of bonuses and penalties, but I could never get the score to
where I wanted it to be without going totally out of control, I felt.
Do you get examples where your program sacrifices pawns for bizarre reasons, which
might be attributable to efforts to be smarter in positions like this?
bruce
Robert Hyatt
brucemo (bru...@nwlink.com) wrote:
I'm more than a little worried about Crafty, because if I set the position up
after the first 5 mvoes as you mentioned, I get an eval of -.030, which means
it sees enough positional compensation to make the pawn down a wash.
For Crafty, the white knight is very good, because it blockades the black pawn,
which makes the two black rooks behind this passer useless (Crafty has code to
handle this case, it still thinks one rook behind it is o.k., but the second
is useless if the pawn is blockaded.) The open file is good, the passed
pawn for black is bad, although it's passed but isolated... In short, I'm
a little concerned that a pawn sac here might be good, but in the general case
it might be a loser. I see crafty fairly frequently sacrifice a pawn to
rip the opponent's pawn structure (he may end up a pawn +, but with 5 or
6 isolated pawns for example), but this looks pretty dangerous to me,
because if a rook or pair of rooks get traded, white suddenly is going to
be embarassed when it might not be able to find the compensation it believes
is there.
I'm studying this, but more with an eye for why this ought to be worse,
rather than why it should be better. Seems to me it's a position that is
a *dangerous* one to solve. Seems particularly dangerous when I see the
+.7 of Rebel8, or the +.4 of Dark Thought... I'd expect to see them lose
a pawn here and there when it doesn't work...
Bob
Ernst A. Heinz
Bruce,
> I know this sounds like an accusation, but I don't mean it to be.
because you are definitely one of the most objective and polite posters on
this group, we would never ``smell'' any such intentions behind your message.
Heck, whenever we met we surely had a good time together, didn't we?! :-)
> Please don't take this the wrong way, but did you guys ever tune for this
> position or did this just happen?
No, we did not specifically tune for this position and, of course, we can
easily change some evaluation parameters in order to make DarkThought/Jakarta
play Ke3, Kf3 etc. Right now, however, we like the current combination of
evaluation parameters best ... :-))
> It's fascinating that there'd be enough compensation so early. I wonder if the
> compensation is evident if you do an eval of the position after
> 1. d5 cxd5 2. e5 R6d7 3. Nd4 ?
Yes, it is. DarkThought/Jakarta evaluates the position after 3. Nd4 as
+0.25 pawns for White.
Given its strong pawns and pieces (supreme knight, dominating rooks) versus
Black's inactive pieces seemingly tied forever to its *very* weak pawns,
this evaluation is not at all unreasonable.
Furthermore, a 16-ply search after 3. Nd4 makes its evaluation climb further.
> Do you get examples where your program sacrifices pawns for bizarre reasons,
> which might be attributable to efforts to be smarter in positions like this?
Well, this seems to be a rhetoric question because it is a tautology for
every non-omniscient (i.e. *any*) static evaluation allowing for positional
compensation of material imbalance.
As stated above, right now we like our current combination of evaluation
parameters best -- but who knows what we prefer tomorrow ... :-))
Michael F. Byrne
> What interests me is not the ply 11 PV, but the ply 6 and 7 PV's. It's
fascinating
> that there'd be enough compensation so early. I wonder if the
compensation is evident
> if you do an eval of the position after 1. d5 cxd5 2. e5 R6d7 3. Nd4 ?
>
> Please don't take this the wrong way, but did you guys ever tune for this
position or
> did this just happen? I know this sounds like an accusation, but I don't
mean it to
> be. I would have no problems tuning for this position since there are a
lot of
> positional aspects here that do deserve a big bonus. I made a
half-hearted attempt to
> tune for this, by trying to analyze some aspects of the position that
deserved
> attention in the form of bonuses and penalties, but I could never get the
score to
> where I wanted it to be without going totally out of control, I felt.
>
> Do you get examples where your program sacrifices pawns for bizarre
reasons, which
> might be attributable to efforts to be smarter in positions like this?
>
> bruce
>
I found this position extremely interesting....and initially so of course I
ran crafty with it
..crafty was close....since i run crafty on ics as fitter..I decided that
in this position i want fitter to play d5
so after some adjustments I got this:
White(1): st
search time set to 600.
White(1): test
=======================================================
! bk.2 !
=======================================================
3r1k2/4npp1/1ppr3p/p6P/P2PPPP1/1NR5/5K2/2R5 w
+---+---+---+---+---+---+---+---+
8 | | | | *R| | *K| | |
+---+---+---+---+---+---+---+---+
7 | | | | | *N| *P| *P| |
+---+---+---+---+---+---+---+---+
6 | | *P| *P| *R| | | | *P|
+---+---+---+---+---+---+---+---+
5 | *P| | | | | | | P |
+---+---+---+---+---+---+---+---+
4 | P | | | P | P | P | P | |
+---+---+---+---+---+---+---+---+
3 | | N | R | | | | | |
+---+---+---+---+---+---+---+---+
2 | | | | | | K | | |
+---+---+---+---+---+---+---+---+
1 | | | R | | | | | |
+---+---+---+---+---+---+---+---+
a b c d e f g h
solution 1. d5 2. exit illegal move.
end-game phase
clearing transposition table
clearing pawn hash tables
time limit 10:00
depth time score variation (1)
5-> 1.23 -0.030 e5 Rd5 Ra1 f6 Rac1
6 1.58 -0.010 e5 Rd5 Ke3 b5 axb5 cxb5 Rc7
6-> 2.75 -0.010 e5 Rd5 Ke3 b5 axb5 cxb5 Rc7
7 3.94 -0.020 e5 Rd5 f5 f6 e6 b5 Nxa5 Rxd4 Nxc6
Nxc6 Rxc6 bxa4
7 9.94 ++ d5
7 10.70 0.040 d5 cxd5 e5 R6d7 Nd4 g6 hxg6 Nxg6
7-> 10.78 0.040 d5 cxd5 e5 R6d7 Nd4 g6 hxg6 Nxg6
8 12.61 -0.080 d5 cxd5 e5 R6d7 Nd4 Ra8 Ke3 f6
8 13.15 ++ e5
8 15.54 -0.020 e5 Rd5 f5 f6 e6 b5 Nxa5 Rxd4 Nxc6
Nxc6 Rxc6 bxa4
8 31.78 ++ Ke2
8 36.12 -0.010 Ke2 R6d7 e5 Rd5 Kd3 b5 axb5 cxb5 Rc7
8-> 48.30 -0.010 Ke2 R6d7 e5 Rd5 Kd3 b5 axb5 cxb5 Rc7
9 1:18 -0.070 Ke2 f6 e5 Rd5 exf6 gxf6 Ke3 b5 Nxa5
bxa4 Nxc6 Nxc6 Rxc6
9 1:20 ++ e5
9 1:25 -0.020 e5 Rd5 f5 f6 e6 b5 Nxa5 Rxd4 Nxc6
Nxc6 Rxc6 bxa4
9 1:29 ++ d5
9 1:35 0.080 d5 cxd5 e5 R6d7 Nd4 Ra8 Ke3 g6 hxg6
Nxg6
9-> 2:06 0.080 d5 cxd5 e5 R6d7 Nd4 Ra8 Ke3 g6 hxg6
Nxg6
10 2:23 0.110 d5 cxd5 e5 R6d7 Nd4 Ra8 Ke3 f6 e6
Rd6
10-> 2:30 0.110 d5 cxd5 e5 R6d7 Nd4 Ra8 Ke3 f6 e6
Rd6
time: 2:30 cpu:100% mat:-1 n:4983381 nps:33131
ext-> checks:53858 recaps:15557 pawns:27537 1rep:30763 unexh:20311
nodes full:874961 quiescence:4108420 evals:3473564
endgame tablebase-> probes done: 0 successful: 0
----------------------> solution correct.
test results summary:
total positions searched.......... 1
number right...................... 1
number wrong...................... 0
percentage right.................. 100
percentage wrong.................. 0
total nodes searched.............. 4983381
average search depth.............. 10.0
nodes per second.................. 33100
Extensions:
check ............................ 53858
recaptures........................ 15557
pawn pushes ...................... 27537
one reply ........................ 30763
uneven exchange .................. 20311
correct solution time(seconds) ... 150.4
I then ran the entire wac300 and this modified version of crafty is getting
298/300
correct in 10 minutes....fitter is currently hovering near 2600 ics with
this current version
brucemo
john quill taylor wrote:
> I ran MChess Pro 5.0 on this through completion of 13 ply, for 15 hours,
> and after 1.6 billion nodes searched, with a value of +0.18 it likes:
>
> 1.f5 R6d7 2.Kg3 Rc8 3.d5 Rcd8 4.dxc6 Nxc6 5.Rxc6 Rd3+ 6.Kf4 and so on.
>
> It found 1.f5 immediately, then switched to 1.Kf3 at 56 seconds, but
> then it switched back to 1.f5 permanently at 39 minutes.
Yup. If you discount 1. d5, there is no obvious best move, in my opinion.
bruce
Howard Exner
The following is me taking the white pieces against
Rebel 8.0. I thought that simple transposing to an
endgame might win for white. Computers seem to want
to go after the black b pawn but this example of exchanging
rooks looks good for white. Of course play on both sides
is up in the air for criticism. I would like to take sole credit
for all of my moves but must confess to the odd "takeback"
in order to get to the truth behind this interesting position.
Exner against Rebel 8.0 Score:?
Date:16-11-1996
LEVEL :Tournament:40 in 02:00
1. Nb3-d4
04:59 11.00 0.54 Nb3-d4 Rd8-e8 Rc3-c2 Kf8-g8 Kf2-f3 Kg8-h7
Kf3-g3
1. Rd8-e8
03:16 10.00 -0.37 Rd8-e8 Kf2-g2 g7-g6 h5xg6 Ne7xg6 Kg2-f3
Ng6-e7
2. Kf2-e3 Rd7-d8
00:59 10.00 -0.48 Rd7-d8 Rc1-b1 Rd8-c8 Rc3xc8 Ne7xc8 Rb1-b5
Re8-d8
3. f4-f5 Rd8-c8
01:57 11.00 -0.32 Rd8-c8 Rc3xc8 Re8xc8 Rc1xc8+ Ne7xc8 Nd4-b5
Nc8-e7
4. Rc3-c7 Rc8xc7
01:35 11.00 -0.32 Rc8xc7 Rc1xc7 Re8-c8 Rc7xc8+ Ne7xc8 Nd4-b5
Nc8-e7
5. Rc1xc7
02:24 11.00 0.33 Kg2-f3 Re8-d8 Kf3-e3 Kf8-g8 Ke3-d3 Rd8-d7
Rb1-c1
5. Re8-c8
06:05 12.01 -0.45 Re8-c8 Rc7xc8+ Ne7xc8 Nd4-b5 Kf8-e7 Ke3-d4
Ke7-d7
6. Rc7xc8+ Ne7xc8
00:00 9.00 -0.41 Ne7xc8 Nd4-b5 Nc8-e7 Ke3-d4 Kf8-e8 Nb5-c7+
Ke8-d7
7. Nd4-b5 Kf8-e7
08:16 14.00 -0.55 Kf8-e7 Ke3-d4 Ke7-d7 Kd4xd5 Nc8-e7+ Kd5-e4
Ne7-g8
8. Ke3-d4 Ke7-d7
06:07 14.00 -0.82 Ke7-d7 Kd4xd5 Nc8-e7+ Kd5-e4 Ne7-g8 Ke4-f4
Kd7-c6
9. Kd4xd5 Nc8-e7+
04:49 13.00 -0.93 Nc8-e7+ Kd5-e4 Ne7-c8 Ke4-f4 f7-f6 e5-e6+
Kd7-c6
10. Kd5-e4 Ne7-g8
11:17 14.01 -0.94 Ne7-g8 Ke4-f4 f7-f6 e5-e6+ Kd7-c6 Kf4-e4
Kc6-c5
11. Nb5-c3 Kd7-c6
05:45 14.00 -0.67 Kd7-c6 Ke4-d4 Ng8-e7 Kd4-c4 Kc6-d7 Nc3-d5
Ne7-c6
12. Ke4-d4 f7-f6
00:00 15.00 -1.00 f7-f6 e5-e6 Ng8-e7 Kd4-c4 Kc6-d6 Nc3-e2
Ne7-d5
13. e5-e6 Ng8-e7
04:14 14.00 -0.96 Ng8-e7 Nc3-b5 Ne7-g8 Kd4-c4 Ng8-e7 Nb5-d4+
Kc6-d6
14. Kd4-c4 Kc6-d6
03:19 14.00 -1.08 Kc6-d6 Kc4-b5 Kd6-e5 Kb5xb6 Ke5-f4 Kb6-c5
Kf4xg4
15. Kc4-b5 Ne7-c8
00:00 16.02 -1.36 Ne7-c8 Nc3-b1 Kd6-c7 Kb5-a6 Kc7-c6 Nb1-a3
Nc8-d6
16. e6-e7 Nc8xe7
06:36 16.00 -0.57 Nc8xe7 Kb5xb6 Ne7-c8+ Kb6xa5 Kd6-c5 Nc3-e2
Nc8-d6
17. Kb5xb6 Ne7-c8+
12:00 15.00 -1.29 Ne7-c8+ Kb6xa5 Kd6-e5 Ka5-a6 Nc8-d6 a4-a5
Nd6-c4
18. Kb6xa5 Kd6-c5
05:38 15.00 -0.61 Kd6-c5 Ka5-a6 Kc5-b4 Nc3-d5+ Kb4xa4 Nd5-b6+
Nc8xb6
I stopped the game here considering it a win for white.
Chess program: Ed Schr"der Speed:
133 Mhz 486AMD
Book openings: Jeroen Noomen
Copyright: E.G.H. Schr"der BV
Martin Borriss
In article <328CED...@nwlink.com>,
Yesterday I confronted a world top GM (who is known to Ferret as well ;))
with the BK #2 position. We didn't feel like analyzing; but after a brief
look he basically said the same about the position I said (e.g.,nice
compensation, but probably no forced win).
Martin
Martin Borriss
In article <328CF4...@nwlink.com>,
>> I think that many of those book lines have to be checked very carefully.
>> E.g. Ferret will probably defend such a position as good as say Kmoch.
>> That's why I wanted to see a program win with this sacrifice against another
>> program first.
>
>Here are two games for you. The first is a playoff between my program and itself, at a time control
>of 30 0, on a P6/200, after the moves 1. d5 cxd5 2. e5 R6d7 have been played. The second is the
>same deal only I didn't fake any moves in first (I was very surprised at 1. Rb1, normally it play
>s 1. Ke3 or 1. f5). My apologies for PGN adherents, as I don't know how to specify a FEN string in
>a PGN file, I just guessed.
>
I did not draw any conclusion from them. The first game showed that Ferret
does not like the thematic treatment of the position. So it came close to
losing this as white, I thought at the time.
With the second game there seems to be an error in the PGN file (3. Rd6
looks illegal since the rook is already there). If you correct this, I'll be
glad to look at it.
*If* someone wants to conclude something from it, then it could be that 1.d5
is not much stronger than 1.Rb1 :)
Martin
Vincent Diepeveen
>Martin Borriss wrote:
>
>> Programs which have adjusted their eval based on this position run a huge
>> risk to give pawns in similar positions where they really are not getting
>> much in return. That is why I am satisfied with Gullydeckel playing 1.Ke3
>> here :)
>
>I think this is a fine attitude, although I think that it is also a good
>thing to try to have material and positional terms balanced in such a way
>that a program does attempt such enterprises as 1. d5 occasionally :-)
>
>bruce
Diep very easily sacrafices pawns, but in this ending it definitely
is not going to play d4-d5 already at 6 ply.
after d5,cd e5,R6d7 Nd4
You are 1 pawn down, black is a pawn up. -100
Where in the starting position diep is already very optimistic
about whites chances (more board control and more centre control).
This goes to a point that diep needs over 2 pawns positionally to play
d4-d5, where giving away a pawn in centre also decreases positional score
of white considerably (important pawn), and black gets (although blocked)
a freepawn (little plus). It needs a huge depth to see this.
After 5 ply this is of course
not possible to see except if you adjusted parameters to it and/or give
a bonus for known patterns that occur how surprisingly in this
position.
--
+--------------------------------------+
|| email : vdie...@cs.ruu.nl ||
|| Vincent Diepeveen ||
+======================================+
Robert Hyatt
Martin Borriss (bor...@inf.tu-dresden.de) wrote:
: In article <328CED...@nwlink.com>,
: brucemo <bru...@nwlink.com> writes:
: >Martin Borriss wrote:
: >
: >> Programs which have adjusted their eval based on this position run a huge
: >> risk to give pawns in similar positions where they really are not getting
: >> much in return. That is why I am satisfied with Gullydeckel playing 1.Ke3
: >> here :)
: >
: >I think this is a fine attitude, although I think that it is also a good
: >thing to try to have material and positional terms balanced in such a way
: >that a program does attempt such enterprises as 1. d5 occasionally :-)
:
: Yesterday I confronted a world top GM (who is known to Ferret as well ;))
: with the BK #2 position. We didn't feel like analyzing; but after a brief
: look he basically said the same about the position I said (e.g.,nice
: compensation, but probably no forced win).
:
: Martin
:
: --
: Martin....@inf.tu-dresden.de
I had the exact same conversation with a GM over the weekend myself. My
concern is that "compensation" seems to be temporal in nature, that is, you
get it, and capitalize on it quickly, or else it slowly evaporates over
time. I've already seen Crafty lose a game here or there by sacrificing a
pawn to shred the opponent's position, only to find it really wasn't
shredded enough to justify being a pawn down. I certainly don't want to
see any increase in this activity.
My bottom line is that I won't spend any time trying to get Crafty to solve
this (or similar) positions. It's easy to make it give up pawns, the hard
part is making it prove that this decision was right. Given two computers,
I'd likely rather have the black side of that position, forcing white to
prove the pawn sac was sound. Machines are quite good at defending such
positions and if it can force the trade of a rook or two, that missing
pawn will get much larger. If white has to be concerned with avoiding
such trades, he might get pushed back into a position where the missing
pawn still gets larger...
Jouni Uski
I think that all BK position are 1) too easy or
2) not correct at all!
Jouni
brucemo
The ones that are pure tactics are getting very easy since hardware is getting faster
and people are using the null-move, but some of them are still pretty good.
Mine takes 9 seconds to find the right move in position #10, and 7 seconds to find the
right move in position #22, on a P5/133. Think of how long this would have taken on
an 8086.
bruce
Harald Faber
Hello brucemo,
I think I have an interesting aspect to tell you:
b> This is position #2 from the Bratko-Kopec test suite:
b>
b> 3r1k2/4npp1/1ppr3p/p6P/P2PPPP1/1NR5/5K2/2R5 w - - 0 1
b>
Rebel *7* needs 16 seconds (486/133 256kB hash) to find the solution.
Rebel *8* needs 10 seconds, same hardware, both running under Win95
Doesn't matter the size of the hashtables
b> 1. d5! cxd5
b> 2. e5! R6d7
b> 3. Nd4
That's Rebels main line.
Ciao and see ya
Harald
--