Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Statistics errors in opening books created with the Shredder8/Fritz-GUI

Author: Norm Pollock

Date: 20:09:24 12/28/04

Go up one level in this thread


On December 28, 2004 at 22:10:45, William Penn wrote:

>On December 28, 2004 at 21:25:59, Stephen A. Boak wrote:
>
>>On December 28, 2004 at 20:54:45, William Penn wrote:
>>
>>>Statistics errors in opening books created with Shredder8/Fritz-GUI
>>>-------------------------------------------------------------------
>>>
>>>I have been creating opening books for quite awhile with the Shredder 8
>>>CB/Fritz-GUI. [I'm talking about the same GUI that comes with Fritz 8, Junior 8,
>>>etc. - not the "Classic" Shredder GUI. Too bad such ambiguities exist, but
>>>hopefully that is clear enough.]
>>>
>>>The procedure (from the chessboard window in the GUI) is:
>>>1) File menu, New, Openings book... (creates a new opening book)
>>>2) Edit menu, Openings Book, Import Games... (imports games from any desired
>>>database)
>>>
>>>The resulting opening book can be displayed in the Book Pane. That is available
>>>by default as one of the tabs in the Notation pane, along with the Notation and
>>>Scoresheet tabs. Or it can be displayed separately by selecting "Extra Book
>>>Pane" from the Window menu, Panes...
>>>
>>>For any given position on the chessboard, the opening book shows the number of
>>>games, percentage scored, average rating, performance, and probabilities -
>>>overall, and for each of the possible moves in the position from the database
>>>involved. A histogram (bar graph) can also be displayed at the bottom of the
>>>Book Pane with the number of white wins, draws, black wins, percentages, etc.
>>>
>>>Much to my dismay, I have just discovered that the data and statistics shown are
>>>inaccurate. There must be a bug somewhere.
>>>
>>>While the total number of games in a given position appears to be correct in all
>>>cases, the other statistics & data for the variations is incorrect. For example
>>>there may actually only be one (1) game in the database for a particular move
>>>from the position, but the Book Pane may say that there are 3 games, or 5 games,
>>>etc. with corresponding incorrect statistics. Search of the database involved
>>>for that position proves that the displayed statistics are incorrect, which is
>>>easy to do when only a few games are involved.
>>>
>>>This error always displays more games than actually exist in the database.
>>>Exactly where the "extra" games come from is a mystery, because they do not
>>>exist in the database. It could be some kind of tree-building error when the
>>>games are imported. Or it could be an error in how the data is displayed (more
>>>of a GUI-based error). I don't know which.
>>>
>>>I also don't know if the error relates to the Length designated when importing
>>>games. That dialog box which appears asks you to select an Absolute length or an
>>>ECO-relative length of 1-100. This determines how big the tree is. The larger
>>>the Length selected, then the bigger (more MB or GB) the resulting opening book.
>>>I usually choose a Length in the range of 15-35, depending on the size of the
>>>database.
>>>
>>>I also don't know if the error relates to the size of the database involved.
>>>Most of my databases are large, the largest over 3,000,000 games. Or perhaps
>>>there is some limit on the size of the resulting book files, but I don't know.
>>>
>>>At this point, I also don't know exactly how bad the error is. If it's only a
>>>handful of games, it wouldn't matter very much when the sample of games in a
>>>given position is large.
>>>
>>>The uncertainties involved makes the resulting statistics practically worthless,
>>>pending further research. Too bad, because it could be quite valuable if
>>>accurate based on the database involved.
>>>
>>>At this point, I'm not aware of any way to avoid this bug. Does anyone know how
>>>it can be avoided?
>>>
>>>My apology for saying "I don't know" so often, but alas there is no detailed
>>>documentation or instructions for these operations insofar as I'm aware. You
>>>just have to discover them on your own, then try to get lucky, I guess. Alas
>>>that is typical of Chessbase documentation. They assume you already know
>>>everything.
>>>WP
>>
>>I have detected similar puzzles in the past.
>>
>>Some thoughts (ideas only):
>>
>>1. Transpositions may mean the position arrived at after making the displayed
>>tree move was reached from different prior positions.
>>
>>Thus a search from the current position may only show one instance.
>>
>>2. Some transpositions may be included only in the notes within a game, but not
>>with a separately saved database game entry.
>>
>>3. Reversed color openings may be included in the statistics, but a search may
>>not find them, even if "equivalent".
>>
>>I suppose a test, creating a small database and corresponding tree, *may* show
>>something, but then the problem may be hard to replicate.
>>
>>Let us know if you learn anything more.
>>
>>Regards,
>>--Steve
>
>Thanks for your thoughts. None of my databases include any annotations, only raw
>game scores, which rules out that possible source of error at least. Here is a
>very simple example of how the Book Pane appears for a particular opening book
>and position, which I happen to be studying at the present time:
>
>The Book Pane says there are a total of 2 games in the underlying database for
>the current chessboard position. That is correct.
>
>It also shows there is 1 game in the database after the move 9...a6 is made.
>That is also correct.
>
>Now the error...
>It says there are 5 games in the database after the move 9...d6 is made. That is
>divided into 2 white wins, 1 draw, and 2 black wins. That is very incorrect! In
>fact, there is only 1 game in the database after 9...d6 (a draw). Where could it
>possibly find the 4 extra games, which do not exist in the database from which
>the opening book was prepared?
>WP

Did you have book learning on? Games played with book learning on will adjust
the stats.



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.