Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Creating Opening books ==> don't use CAP data.

Author: Robert Hyatt

Date: 19:03:26 07/06/00

Go up one level in this thread


On July 06, 2000 at 19:41:31, Vincent Diepeveen wrote:

>Hello,
>
>this is a post i just did at crafty list after a long number of mails about
>a project called CAP which people tend to believe they can use it for
>creating their openingsbook. Some warnings are on their place, when you
>want to let such a book compete with commercial books.
>
>I first showed that CAP goes completely wrong already quite soon in opening.
>Like 1.d4,d5 2.c4,e5 3.dxe5 there best move is d4 all other moves you can
>go home and rest but first resign the game.

This is a bogus title for a thread.  CAP isn't right in all cases.  Neither
is "most popular move".  Nor "move suggested by Kasparov (or Anand or Karpov
or any other GM)".

CAP data _is_ good and reliable.  You simply have to use it as one of several
decision-making ideas.  Not _the_ decision making idea.  Weighted with other
things like frequency, learning, even simple positional evaluation, CAP data
can contribute to a better move choice.  Used by itself, it can also contribute
to a bad choice.  As can even a top GM's suggestion(s).

If you don't want to use the data, don't.  But I don't see any reason to
suggest that others should not try it.  They might have a better idea than
you do (or than I do).

Dismissing something just because you don't see (yet) how to use it is not
a good way to make progress.  Had I dismissed bitmaps so quickly, I would have
missed something that has turned out to be very interesting and useful.




IMHO...





>--------------------------------------------------------------
>
>In my previous examples you might argue that the positions are of
>material inequality. But now please check cap scores for the next
>opening:
>
>1.e4,e5 2.Nf3,Nc6 3.Bb5,a6 4.Ba4,Nf6 5.O-O,b5
>
>I didn't check yet. I just give a mainline in the Spanish here.
>
>I remember that at world championships Stefan-Meyer
>Kahlen fiddled at my laptop together with Rudolf Huber and Thorsten Czub
>with DIEP. Their conclusion was that DIEP was way better for CAP project
>as crafty, but i didn't understand what they meant, as i had never
>taken CAP seriously. Now i realize still some people are busy with
>it. I'll explain in this posts why i find a single computer analysis
>of a line not very smart to base a book upon.
>
>Now keep into mind that all busts of my book at tournaments were busts
>where CAP would have gone wrong sooner in the opening or at the same
>point.
>
>However, the problem is more complex than just that
>it's also that programs are quite strong in winning games compared to
>their real strength.
>
>As Bob already indicated if a line A is played a million times and
>a line B only a 100 times, then obviously you take line A, yet also that
>is not entirely true. Line A might be refuted or a move of the thumb also.
>
>When starting with the first: making your move based upon the outcome of
>a few minutes of search of a program:
>Programs are not near the human
>strength, but they just don't blunder, so in contradiction to me
>computers don't have a weak point, but when there is no tactics in
>a position then a computer is not a part of my strength. Remember i'm
>just 2255 FIDE, getting to 2400 soon perhaps and hoping to get my IM
>within a couple of years, but i'm not near the incredible brilliant level
>of the 2600+ players, who also have the habit to find new moves behind the
>tournament board and the majority of new moves doesn't get found at home
>as most people do you want to believe. Basically Kasparov does and a few
>correspondence players do.
>
>Shredder who clearly is world
>champion past few years, i can't say that i'm very impressed about how it
>positionally plays, but it just doesn't lose quick. Shredder is extremely
>bad short after opening. Classical preprocessor problem. If we look at
>games from the world championships like the Shredder-Ferret tie then we
>clearly see already short after opening the Rfe1 manoeuvre where Shredder
>puts itself in a pin, something no human would ever do. Yet it's not
>tactical losing directly. To lose black must have a strategic plan which it
>developed rather late (perhaps too late). So obviously all programs
>suffer from not having the same human insight. But let's continue to
>describe how i feel about my book versus CAP.
>
>Considering the more or less automatic way in which i constructed
>book about a year ago short after the world champs,
>let's see where my program went wrong in tournaments.
>
>DIEP's main book, and numerous crafty clone books have been constructed,
>in an automatic way based upon the number of games played. Added to that
>i have a small hand tailored book which was really small, so always
>a transition from tournament book to automatic generated book takes place.
>I'm not happy about that transition. Not at all, see below why.
>
>The past 8 months:
>
>  - dutch open diep played 4 games against commercial programs,
>     - chesstiger
>     - the king
>     - quest
>     - nimzo
>
>    Let's chronologically handle this. Consider that diep has a
>    very small tournament book, not worth mentionning. These programs
>    have thousands of hand made book moves. Not generated automatically
>    whatsoever.
>
>    - against chesstiger diep gets completely busted out of book
>      directly after book a pawn goes even and a lost position gets to
>      board with a pawn less. I still don't understand how DIEP managed
>      to draw that. that's not important here. important is to realize
>      i get completely lost out of book against a chess program and that
>      because of that the opponent can no longer lose, despite incredible
>      pathetic strategically and especially positoinal play caused by
>      preprocessing of Tiger. How can you draw a game with a pawn
>      up in that position? Incredible! But still... ...a good job in the
>      opening and it could no longer lose!
>    - against The King, there The King gets great play after opening,
>      not exactly huge advantage, but simply easy play,
>      only because of chosing the wrong plan and the fact that i ran
>      at a quad saves diep's ass to an equal position which later
>      diep manages to convert to a pawn more in endgame which still remains
>      a draw. So good tactics was enough to keep it a draw. Also the fact
>      that endgame of both programs wasn't at high level kept it a draw,
>      as any human would have seen that just winning the pawn wasn't important
>      where diep just got that pawn and only realized years later it was a
>      draw.
>    - against Nimzo diep gets a mainline at board. Suddenly Nimzo plays
>      a different move somewhere. The move is bad, but the automatically
>      generated book completely mistakes there. I had 20 games in book.
>      19 games or so a3 was played with a reasonably good chance for winning
>      with that line, according to statistics. Just 1 game a different move
>      was played, which got ignored of course 1 game is not convincing. 19
>      games are however.
>      So diep plays a3 and is already having a lost position, a3 is a good
>      answer to the mainline move, but a side line move was played, lots of
>      2000+ players probably have made the same mistake. 19 to be exact.
>      Still theoretical
>      humans consider that as still a bit unclear as white has tactical
>      pressure on black. Of course computer doesn't fall for the pressure
>      so we can consider it as a dead lost position. The bad thing is
>      that both programs didn't see that black was won, only when it was
>      far too late both programs realized how dead lost it was for white.
>      Afterwards i can consider that as a chanceless loss of an automatically
>      generated book. So this was seemingly a good game from Nimzo, as
>      outsiders might say : "from a negative score of -1.0 it fought its
>      way back", but fact is that programs always underestimate positions
>      with 2 passed pawns that start running. Only when they get to 6th row
>      programs realize suddenly how strong they are.
>    - against Quest was even more horrible. Diep plays a move (b4??) which
>      was played in theory a lot by grandmasters, however nowadays line is
>      completely refuted. It's a ++-- case. +10.0 out of book for Quest.
>
>  Against Quest also CAP data would not have saved DIEP's ass. So
>  i would not have won this tournament even with CAP data which makes
>  CAP data useless. Only way to do better against these programs next year
>  is to have an up to date book.
>
>  It's obvious that Nimzo book is especially made to do well against
>  automatically generated books. Note that Nimzo book is the same as fritz6
>  book. Nimzo 7.32 book is only a bit newer than that i can assume.
>
>Now you convinced that making a hand made book is necessary?
>
>Horror strikes at spanish champs for DIEP at christmas 99.
>
>Diep had won all games at Spain champ. Tiger had lost one.
>Just a draw was enough to win the title for DIEP.
>
>I had prepared a line at home to play against Tiger both with white
>and black. I shipped
>before the champs that book to Spain, and hoped this line would get
>on the board.
>
>EXACTLY the line came on the board. However let's talk about the
>openingsbook line. I never have played that opening, i just had written
>down the line from a book, i had not even seen a board when preparing
>that line, book said += in that position.
>
>However DIEP lost it chanceless against Tiger!
>
>Note that Tiger played
>that game quite well, but still, i clearly did a bad job in the
>openings book, because what happened?
>
>Post mortem analysis showed that my openingsbook was simply 15 years too old.
>My book was out of 1984. Sincethen grandmasters concluded that this line
>though tactical giving much pressure at black (tigers side) objectively
>seen the line is losing for white, so someone who doesn't give away
>pieces (for example computers or grandmasters) will win it with black
>in a chanceless way. In new theory books the side line
>played by DIEP is not even considered and given a big questionmark in
>the notes.
>
>So making hand tailored books is nowadays not even enough.
>No what you need is a book at the same level as i play chess myself,
>and i'm tactical not as near as good as my program!
>
>I'm positionally/strategically way above it, but ones or twice a game
>i make a big blunder. Usually 1 blunder a game (if i lose the game).
>
>Computer never makes big blunders, only bad moves in openings or
>strategic mistake somewhere. That's about it.
>
>So basically you not only need hand tailored book, but also refuted lines
>you need to leave out of book and you need to take into account that
>a program just needs an OBJECTIVE correct position. Too many grandmaster
>take into account that mankind makes tactical mistakes. So also the
>GM analyzes should be taken with a bit of salt.
>
>Obviously Nimzo 7.32 book is taking this all into account.
>
>Have you ever seen a book loss of Nimzo, or from Fritz at the different
>championships?
>
>No? you haven't? How comes, they used CAP? I'm sure they didn't.
>It's all based upon human analyzes and theorem, interpreted by a strong
>playing human and based upon thousands of auto232 games and all together
>it's years of work.
>
>So where you still need to realize that you need a handtailored book,
>i'm already a step further and i recently bought the latest theory books
>in order to keep up to date with the theory, because hand tailored books
>also lose if you have outdated theory.
>
>This where those automatically generated books already don't take the
>date of games into account. If they would, they still depend upon
>the AVERAGE move played. I can assure you: the average move in this
>world is a 1500 rated move. Programs are stronger as 1500. But programs
>are in opening not as strong as strong GMs are. They're not even
>near FM level in opening.
>
>Now hand tailored books are at least GM level books, So programs searching
>for openingsbook moves gotta be way less than that.
>
>To order books in magnitude of strength
>            A  - CAPdata/automatically generated books
>
>            <big gap>
>
>            B - hand tailored books based upon GM commentary
>
>            C - hand tailored books taken into account latest theory
>            D - hand tailored books taken into account latest theory and
>                objective judgement of a position
>            E - same as D but now with years of testing and book updating
>                and especially taking into account how a program reacts
>                on the booklines based upon real games as played by thousands
>                at home at tournament level (so not blitz level).
>
>I'm trying to make the step from A with a bit of B
>directly to C, and i'm still one step behind some of
>the real great books.
>
>But the CAP data is still at A without a chance that its
>holy believers will ever get to B, because seemingly they've
>got nothing to do.
>
>Vincent Diepeveen
>diep@xs4all.nl
>
>----
>to unsubscribe, send email to majordomo@cis.uab.edu.  In the body of
>the email, enter "unsubscribe crafty-list" (without the quotes).



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.