Author: Albert Silver
Date: 13:38:30 01/08/00
Go up one level in this thread
On January 08, 2000 at 09:56:37, Robert Hyatt wrote:
>On January 08, 2000 at 00:37:32, John Warfield wrote:
>
>>On January 07, 2000 at 23:42:00, Robert Hyatt wrote:
>>
>>>On January 07, 2000 at 19:18:10, Chris Carson wrote:
>>>
>>>>On January 07, 2000 at 18:16:44, Robert Hyatt wrote:
>>>>
>>>>>On January 07, 2000 at 13:59:37, Chris Carson wrote:
>>>>>
>>>>>>Bob,
>>>>>>
>>>>>>I know where you stand. :) 2450 to 2500 is a good number,
>>>>>>I completly agree with you and you have posted excellant
>>>>>>arguments and facts to support it (both now and in the past). :)
>>>>>>
>>>>>>I was not sure if Albert had a different line of reasoning. I
>>>>>>respect his views and was just curious. :)
>>>>>>
>>>>>>Best Regards,
>>>>>>Chris Carson
>>>>>
>>
>>
>>
>>>>>
>>>>>The issue is 'what will a 2600 player do and what won't he do'. Today I
>>>>>watched Tiger vs Crafty where the game was very equal going into the endgame
>>>>>with several pieces left, the only difference is that crafty had a pawn majority
>>>>>that would turn into an outside (distant) passed pawn. The opponent gladly let
>>>>>Crafty trade into an ending that was absolutely dead won for Crafty. A 2600
>>>>>player simply would _not_ do that, because it is a trivially won ending...
>>>>>
>>>>>That is the kind of thing a 2600 player won't do... and that is something that
>>>>>a 2600 player _will_ take advantage of, over and over, when he spots it.
>>>>>
>>>>>And that isn't picking on Tiger, for any reason other than it is the one at the
>>>>>top of the SSDF with a rating near 2700. You won't find _any_ 2700 player that
>>>>>will make a basic mistake like that. But you will find _lots_ of computers that
>>>>>can and will fall into it.
>>>>
>>>>Bob,
>>>>
>>>>IMHO:
>>>>2600 players make mistakes, 2700 players make mistakes, 2800 players
>>>>make mistakes (I think GK made some against DB). Sometimes stupid ones.
>>>>I have never seen any person play error free for extended periods of
>>>>time (person or machine). I am not skilled enough as
>>>>a chess player to point those out, but the colums of chess mags/books
>>>>are littered with them. If I get a chance to locate a mistake by
>>>>a top 50 player, I will post it (perhaps a GK vs DB, either match). :)
>>>>
>>>>I even think GK complained about not having access to DB games (the IBM
>>>>team saw this as an advantage in match play) to prepare with. GK thought
>>>>that preparation against the machine would help his performance MPR or
>>>>TPR or PR. :)
>>>>
>>>>I agree with your statements though. :) IMHO: Programs are likley to
>>>>make the same strategic error over and over. Learning solves some of this,
>>>>but not all. This is a major challenge for programmers and programs
>>>>and a useful tool for people to use against machines.
>>>>Adaptability is a huge hole in programs (among others). :)
>>>>
>>>>Am I adding anything here or just being stubborn? If I am not
>>>>adding anything, tell me and I will drop it. :)
>>>>
>>>>Best Regards,
>>>>Chris Carson
>>>
>>>
>>>Two points. I agree about the adaptability problem. It is a huge problem.
>>>
>>>As far as 2700 players making mistakes, I agree. But not terribly gross
>>>ones (very often). _NO_ GM would ever trade into a dead lost king and pawn
>>>ending, for example.
>>
>> Yes, But there are many mistakes that human 2700 players make which programs
>>don't make. The whole point is that computers posses such extrodinary tactics,
>>that many of the weakness are covered and overshadowed by their main strength.I
>>think your being extremely one sided, you carefully highlight all the weakness
>>of computers and overlook the many weakness of humans.
>>
>
>
>Computers do _some_ tactics great. But I have seen GM players take them to the
>woodshed on tactics as well, as at times, the computer simply doesn't/can't go
>deep enough to see the _real_ answer, and the 'phantom answer' it sees can be
>wrong.
>
>But the problem is that the positional holes are significant enough that it is
>possible to exploit them without much risk, because many programs don't struggle
>to keep the game position in a state that favors the computer. While the GMs
>can definitely steer the game into positions that do not favor the machine, if
>it passively allows this to happen.
>
>Again, this comes from watching on ICC. If you'd like to see how bad your
>favorite program can play there, try this: log on, and accept _every_ match
>request from humans rated (say) 2500 and up. Even if they want to play 50 games
>in a row. And watch what happens after a while when they find a weakness they
>can pick on repeatedly. The weakness can be anything, from a feature that is
>not evaluated, to a book hole. And the humans _talk_. So when one finds a
>weakness, you can expect it to be hit on by several players...
>
>And you are correct that 2700 players will occasionally make a mistake that
>a computer wouldn't... but on balance, the computer has a lot more holes in
>its armor than that 2700 player. And when you move the time control up to
>40/2hrs, the tactical errors by the GMs go way down, while the positional
>errors by the computer are unchanged...
>
>
>
>
>>
>> They might underestimate an attack, or overlook an odd
>>>mate in 2 or 3, but they will not consistently trade into a lost ending.
>>>
>>>That is the difference, and the problem. One sort of problem: king and
>>>pawns and a couple of pieces vs king and pawns and a couple of pieces. The
>>>computer has pawns at (say) f7,g7 and h6, the human has pawns at (say) f4, g3
>>>and h3. The computer plays g5. And creates a probably lost pawn ending
>>>(after fxg5, hxg5, white can make a passed pawn on the h-file, while blacks
>>>will be on the f-file, and with no pieces, white probably wins. A GM would
>>>simply not play g5, or else it would not trade pieces after doing so, knowing
>>>that the pieces can restrain the passer and maybe have a chance.
>>>
>>>It is the _kind_ of error. GMs overlook oddball tactics, to be sure. But they
>>>certainly don't overlook simple lost endgame positions. The computer, on the
>>>other hand, won't often overlook simple tactics, but some will blow the endgame
>>>above over and over and over.
It's more than that. Most GMs (always risky to say all) dominate an aspect of
the game which at this point is completely beyond the scope of PC programs: the
transition from middlegame to the endgame. When I say this I don't have in mind
the simple not swapping the queens as the pawn endgame is lost, but something
similar but far far deeper. Sometimes you will see GMs avoid a pawn move or
swapping a particular piece with a board full because of the endgame that might
ensue. The ability to see these things is only due to their deep knowledge of
the endgame and no calculating will make up for it. Sometimes one reads comments
that may seem to border on the mystical unless one is partisan to the same
knowledge as the annotator. These skills are also pretty much non-existent in
players below IM level which is one of the reasons I have argued that the only
way to see how a program will do against a GM is to play it against him. Beating
2300 players 20-1 won't do the trick. It may be a GM-type performance (over
2500) but it still won't say how it will do against GMs, and getting an extra
ply or two won't solve the problem. That's one of the reasons (with all due
respect) when I read someone say that knowledge can simply be compensated by
more calculation, I can only presume they speak in ignorance and do not know
what they are talking about. That's also why I don't see how telling me one
program beats the other programs makes it a GM. There are conceptual tactical
positions that though fairly simple to solve are also beyond the scope of the
present programs. Below is one which took me a few minutes to solve, yet the
programs choke big-time. The reason is because the solution is quite linear and
the final pattern is fairly easy to see, but the total solution requires too
many plies.
Bagirov-Brodsky, 1992
[D]rq1n1r1k/pb1R1ppp/1p2p3/4P3/5BQ1/2P3P1/P4PBP/4R1K1 w - -
The solution begins 1.Bh6!! gxh6 2.Qh4 Bxg2 3.Qf6+ Kg8 4.Kxg2 Qc8 5.Re4 h5
6.Rdd4 Re8 yet even after feeding these moves to the program
[D]r1qnr1k1/p4p1p/1p2pQ2/4P2p/3RR3/2P3P1/P4PKP/8 w - -
see how long it takes to see 7.Rg4+! is winning, followed by 7...hxg4 8.Rxg4+
Kf8 9.Rh4! and Rxh7-h8 mate! (I'll be curious to see if the diagrams come out
right... :-) ). If you know the mating pattern then the solution isn't all that
difficult, just a few pitfalls to avoid while calculating it.
Albert Silver
This page took 0.01 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.