Author: martin fierz
Date: 14:32:09 09/16/02
Go up one level in this thread
On September 16, 2002 at 13:59:02, Dieter Buerssner wrote:
>On September 15, 2002 at 20:25:46, martin fierz wrote:
>
>>> 8/8/4K1p1/4PpB1/3p1P1P/5n2/4k3/8 b - -
>>
>>hi dieter,
>>
>>funny position! the reason your engine fails here is simple: computer eval
>>functions are incredibly dumb when it comes to winning positions. you count
>>material, which normally is good, but when it comes to winning a winning
>>position, this is just a bad idea.
>>when a human master is winning, his only remaining aim is to get into a simple
>>technical win. computers continue to count material. their eval functions prefer
>>a pawn endgame a pawn down to a rook ending two pawns down, even if the side
>>with two pawns less has some counterplay. human masters don't count the
>>material, instead, they think about winning probability. how likely is it that i
>>will win a pawn ending with a clear extra pawn? very close to 100%. how likely
>>is it that i win a rook ending with two extra pawns against an active rook?
>>likely, but not as close to 100%.
>
>I agree, with almost everything. However, I think, that typically the method of
>the chess engine works as well. Indeed, I am partitially using ideas you
>mention, but your input may motivate to do this better.
good for you!
i also agree that mostly the "count material" thing works. but there are
exceptions, and if you study games of human masters you will see that they
prefer to give the opponent no counterplay instead of winning more material.
this is of course in part due to the fact that humans are more likely to make
mistakes in the counterplay scenario, but still.... i remember a game i played
as a 12-year old against a girl where i had already won two pawns and went on to
play something like Qxb7xRa8 and lost because she was attacking my king and i
lost too much time winning that rook. losing against a girl hurts of course, so
i learned my lesson well :-)
i also remember how my chess coach, an FM with a 2350 rating told me that
winning that rook was a really bad idea - whether it was good or not. i should
just not allow any counterplay if i was winning already.
i remember seeing a position here from a game where one side, i think it was
shredder, was totally winning but lost in the end because the other side had two
dangerous connected passed pawns, perhaps a year ago. that was somehow similar
in that the winning side could have gone for a simpler win without the
counterplay but didnt. i also remember recently there were those lovely rook
ending test positions (i think of sune larsson), and a comment by bob totally
shocked me: he was arguing about a certain position where crafty was trading a
rook into a dead lost pawn ending that "the rook ending is lost too". even if
this is true, it doesnt matter at all. the only chance is to keep the rooks on,
because the pawn ending is guaranteed to be hopeless. in my "human eval",
perhaps both have 99% winning chance, but the uncertainty in the rook ending is
much larger.
i think there was an attempt by berliner to make a search algorithm which
thought about evaluations and uncertainties in the evaluations too, perhaps it
was called B*. it never took off though AFAIK, but it's an interesting concept.
>>here's where yace goes wrong in this position: it looks at the winning
>>continuation d3 Kf7 d2 e6 d1Q e7 Qd7 Kf8 Nd4 e8Q+ Ne6+ Qxe6+ Qxe6.
>
>I think, in this position it still is a bit more complicated, because of Bf6
>instead of Kf8, to make Nd4 impossible. I think the mainline goes like this:
>
>d3 Kf7 d2 e6 d1Q e7 Qd7 Bf6 Ke3 Kf8 Kxf4 e8Q Dxe8 Kxe8 and win the h4 pawn (for
>ex. Kg4 Kf7 Nxh4). Now black has knight +2P vs. B, and will win. I am not
>totally sure, that it is the only winning path.
you are right about bf6 of course. i thought about it but you hadnt mentioned it
in your previous post and so i just followed your mainline. this is not the only
winning path though, but it is very convincing. what is convincing about it is
that if i see this line and the final position after Nxh4, then my evaluation is
winning probability 100% - no uncertainty. there are no branches in this line
either, except for Bf6 or not Bf6, but both end up in hopeless positions. if i
happen to see this line in a tournament game, i will not check anything else and
play it.
But what chess engines may do:
>Don't move Kxf4 fast and do some useless Q moves. Then, an engine that counts
>the first repetion as draw, will not move back to d7 with the Q again (there
>seems to be no way, to get the e7 pawn, without saccing the Q). So this
>basically wins a P less. Black must hope, that he can win with one pawn more
>(KQNPPKQBP or KNPPKBP). Lines I tried, yielded in draw. BTW. In very fast game,
>an engine may even spoil this to a loss - it may at an inconvenient point see,
>that a repetion will arise with "normal" means, and sac the N ("We'll, I still
>have a Q vs. B"). Done "clever" enough, and black may win.
>
>>i'm sure it
>>sees this line very quickly - it's only 9ply + a few checks/captures which you
>>probably extend. now it counts: Q-B: +6. note that white has zero compensation
>>here - not even a passed pawn anywhere, no possibility to attack any black
>>piece, this is a dead loss.
>>then it looks at it's game continuation, and gets into that drawn ending, and it
>>counts: Q-P: +8. perhaps, you give the pawn a bonus of 1, because it's on the
>>7th rank. +7. and you play Nxg5??... of course, you are "unlucky" not to win
>>this: a queen against an e-pawn alone would win easily, because of Qf6+ at some
>>point, but you don't have this thanks to white's "lucky pawn" on g5. so how do
>>you fix this?
>>when you are *totally* winning, say +5 or more, check if the losing side has
>>some form of compensation. like a passed pawn. or worse, a passed pawn on the
>>7th rank. then give that a *big* score - much bigger than usual. if you give a
>>passed pawn on the 7th a +3, yace will solve this position instantly... i don't
>>think there are any harmful side-effects, but maybe i'm overlooking something.
>
>So, with the line above, probably +3 for the pawn on 7th will not be enough. But
>your idea is of course valid and very interesting. So, leaving other factors
>out: Yace evaluates the KNPPKB fast slightly lower than +3. The material balance
>in the SxB lines is +8. I think, it would be risky, to give a positional score
>of over 5 for a passed pawn (when you have no other reasons). It may well make
>the engine draw in other positions, that would have been won.
i agree that i wouldnt feel too good about giving huge scores, like 5 in a +8
position. you could also try the other way round: my eval of the KNPP-KB
position is "100% win", not +3. you could try to fix that one alternatively to
fixing the position with the pawn on the 7th.
>
>Another thing came to my mind. I was watching that game, and was chatting with
>the operator. At some point (slightly earlier in the game) I said more or less
>"now f4 is dead and an easy win". Why didn't Yace do this? I extend
>significantly for trading into pawn endgames. So, this may have "helped" to see
>fast, that the d-pawn will promote fast. Of course, without trading into pawn
>endgames, this would have seen too, but with relatively less depth behind it.
>Now, this was in a phase of the game, where the score was still rising - so the
>deeper line going through the pawn endgame, would have gained more score due to
>the deeper depth.
>
>I hope my analyisis is correct. If yes, I wonder, how fast club level chess
>players see all this. Is there another path to win (Besides delaying Kxf4 with
>some Q moves)? Can black take f4 with Q somewhere different as d7?
unfortunately i can't tell you how fast i would see this because i already saw
the lines... all i can tell you is that i looked at the first position and my
instinct, in a few seconds, was to play 1...d3 2.Kf7 Nd4 3.e6 Nxe6 -+. this line
alone gives me enough reason not to look at Nxg5, because i can sac my minor
piece against your passed pawn, and you can't. i would still calculate 4.Kxe6 d2
5.Kf6 d1Q 6.Kxg6 and think about this for a while. but once i see that after Qd7
7. h5 is not possible because of Qe8+ i would play this line and not look back.
therefore i doubt that i would have seen the other line.
if for some reason i was not sure about the line above, or had lots of time to
spend checking, and found the other line with KNPP-KB, i would prefer that line
to the one i just mentioned. although i would be pretty sure that the QP-QBPP is
totally winning, in the KNPP-KB i'd be 100% certain.
my guess is that any player of my level will see one of the two winning lines
fast, in a few seconds. humans have to check their analysis though, so i'd guess
that a typical player of my strength would play one of the two correct lines
after a minute or two of thought.
aloha
martin
>Thanks for your interesting thoughts.
>Dieter
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.