Hi!
Except the pretty technical discussion that follows, those
who just like chess might go to the end, copy the PGN
notation to clipboard and then go Tools / Import one PGN
game and enjoy. ;)
As I just stumbled upon this very nice study below, I also
found it is a demonstration why annotating backwards really
leads to a more correct evaluation of the positions. We had
this discussion recently, but this example in particular
seems a very strong usecase, as the effect is very large.
(An evaluation difference between forward and backward of
0.01 might not show it that strongly.
Starting out at
6N1/3P3k/n2K3p/2p3n1/1b6/2p1p1P1/8/3B4 w - - 0 1
Given the initial position, Shredder 11 evaluates it with -3.5
at depth 16, while Rybka 2.2n (the free one) gives even
-4.3 at this level. Their best line is, and they agree on this,
1.Nf6+ Kg7 2.Nh5+ Kg6 3.Nf4+ Kf5 4.Nd5 Ke4 5.Bh5 Ba5 6.Nxc3+
Note: the study is white to move an WIN. The evaluation at
this point is, however, 3.5 in favour of Black! The initial
moves of this line are correct, both engines fail to find
the key move 3.Bc3+!!. Lets point them with the nose to that
one. After 3.Bc3+ I get on depth 16 from Shredder
-9.99 3.... Kxh5 4.Bd1+ Kg6 5.Ke5 Nf7+ 6.Kf4 Nb8 7.d8=B Nxd8
8.Kxe3 Kg5 9.Bc2 Kg4 10.Kf2 c4 11.Bd1+ Kh3 12.Bc2 Kg4
13.Bd1+ Kh3 14.Bc2 Kg4
and Rybka tells me something like
-9.76 3.... Kxh5 4.Bd1+
So we're now at 9.x in favour of Black. Still, White has to
win...
Both don't like the Queen sacrifice 4.d8=Q. Again pointing
them with the nose to it... After this move the two differ
quite a bit.
Shredder:
-13.34 4.... Nf7+ 5.Ke6 Nxd8+ 6.Kf5 e2 7.Be4 e1=N 8.Bd5 c2 9.Bc4 c1=N
Rybka:
+4.69 4.... Kg4 5.Kc6 Kxg3 6.Qd6+ Kg4 7.Qd3 Kf3 8.Bd1+
Funny enough, Shredder follows the solution right dow to
9...c1=N but gives 13.3 in favour of Black missing Bc4 at
this level. Jumping to that move he sees the mate in 5
immediately, of course. Rybka sees a different line
altogether giving a winning advantage for White here
already, but does not see a mate or whatever. On the next
move they'll find it both.
That's for the forward story.
Going backward through this gives you a quite different
picture.
Shredder e.g. sees already a M11 at move 4..Nf7+. He also
comes to more agreement with Rybka on the evaluation of
4.d8=Q giving a +2.6 here (instead of -13.34!) So, going
backward gained some 16 evaluation points in the correct
direction! Additionally, if you go to the very beginning
given his knowledge from going backward through the game he
now gives a positive evaluation to the initial position.
That is instead of a crushing advantage for Black you now
get almost the same for White as Shredder knows about the
key points from his hash. I also found that Shredder now
will never evaluate to an advantage for Black but always
have advantage for White or even Mate in xx moves.
Additionally, one can notice that the depth of calculation
reached is almost all the time > 16 going backwards as the
engine can reuse previous evaluations. ("Hash usage",
however, does not show 100% all the time.)
You'll observe most of this with Rybka the same way. Rybka
doesn't even really start calculating till move 4 anyway, it
just has its hash and is done. I think Shredder calculates
as it only seems to store the best line in hash.
I didn't try all my engines but as far as I can see Fruit
e.g. shows the same behaviour. It just seems one can fool
Toga. Ie. this engine does not show the above. I don't know
the reason, but maybe it cleans his hashes all the time(?)
This I'd consider a bug in the engine, however.
I did the above on Linux and emulated forward analysis by
hand waiting till the engine reachted depth 16. This took a
lot more time in forward analysis then in backward analysis,
but due to the manual procedure I've not timings for this.
Here is the game in question.
-------------------------- 8< ----------------------------------------
[Event "Twilight Zone 03/2009"]
[Site "Schachzeitung"]
[Date "????.??.??"]
[Round "?"]
[White "?"]
[Black "?"]
[Result "1-0"]
[Creator "G. van Breukelen"]
[Ref "Schachzeitung 03/2009, p.36"]
[FEN "6N1/3P3k/n2K3p/2p3n1/1b6/2p1p1P1/8/3B4 w - - 0 1"]
{White to move an win!}
1.Nf6+ Kg7 $1 ( 1...Kg6 $2 2.Bc2+ Kg7 ( 2...Kxf6 3.d8=Q+ $20 ) 3.Nh5+ Kf7
4.d8=Q $20 ) 2.Nh5+ Kg6 ( 2...Kh7 3.Bc2+ $20 ) ( 2...Kf7 3.d8=Q $18 ) 3.
Bc2+ $3 ( 3.Nf4+ Kf5 $19 ) 3...Kxh5 4.d8=Q Nf7+ 5.Ke6 Nxd8+ 6.Kf5 e2 7.Be4
e1=N $1 8.Bd5 $1 c2 9.Bc4 c1=N 10.Bb5 Nc7 11.Ba4 1-0
-------------------------- 8< ----------------------------------------
--
Kind regards, / War is Peace.
| Freedom is Slavery.
Alexander Wagner | Ignorance is Strength.
|
| Theory : G. Orwell, "1984"
/ In practice: USA, since 2001
------------------------------------------------------------------------------
This SF.net email is sponsored by:
High Quality Requirements in a Collaborative Environment.
Download a free trial of Rational Requirements Composer Now!
http://p.sf.net/sfu/www-ibm-com
_______________________________________________
Scid-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scid-users
