Reply
 
LinkBack Thread Tools Display Modes
  #1   Report Post  
Old January 6th 04, 05:12 AM
Dr. David Kirkby
 
Posts: n/a
Default Skewing results by use of one set of GM games.

I've been thinking about something and would appreciate the comments
of others.

Most chess programs use an opening book, which one will typically
based on a set of GM games. My understanding is that the chess
analysis program whilst 'in book' will evaluate a position based on
the frequency of occurance of the moves from this list of GM games,
since you have used that to create your opening book.

So if the opening move played by most grandmaster games is d4, then
the chess program will give the highest score if white plays d4,
rather than says a4, which would be a rare opening move for a
grandmaster. If the most common reply by black is Nf6, and one plays
that in a game, the chess engine will think that is black's best
response. Is my understanding of this correct?

If one then uses a set of GM games as a 'reference database' in a
database program (chessbase, scid etc) to evaluate the best opening
move, then if the same set of GM games is used, it will of course have
d4 as the most common opening move. Then when you evaluate the
position with a chess engine, it will give the highest score to white
if he plays d4. Likewise the database will show Nf6 as blacks most
common (and so one might infer best) response.

This suggests to me that the results of computer analysis of positions
get skewed if the ***same*** set of GM games are used both to create
an opening book in a chess analysis program AND as a reference
database in scid, chessbase or whatever.

Clearly with the first few opening moves it makes no difference if the
list of GM games is reasonably long. But as you progress further into
the game, the skewing of results might become significant.

Statistics never was my strongest subject, so I'd hate to guess when
this becomes significant, but I can't help feeling one develops a
false sense of what is the best move if both the chess engine and the
database program use the same set of GM games.

I'm using a set of about 27,000 GM games downloaded from the crafy ftp
site to create an opening book and also the same set in scid as a
reference database. I can't help feeling that this is not very
sensible. It seemed reasonably logical at first to use a set of high
quality games for both the opening book and the reference database,
but now I'm not so sure.

Comments ??

Dr. David Kirkby.
  #2   Report Post  
Old January 6th 04, 07:52 AM
Derek Wildstar
 
Posts: n/a
Default Skewing results by use of one set of GM games.


"Dr. David Kirkby" m wrote
in message m...

If the most common reply by black is Nf6, and one plays
that in a game, the chess engine will think that is black's best
response. Is my understanding of this correct?


Not exactly, Frequency of appearance is just one criteria of move
applicability in book.
Here is an example, from Rookie 2.0, of how opening moves are weighted:

"If n moves are available to be played then the probability for a move m
(leading to p) to get chosen is proportional to P(m) = (1-Wp) x Gp x
(1-Rp). This way we favor moves with higher winning expectations and also we
prefer to choose moves that will keep us in the book as long as possible."

Backing up for variable defintions:

Each position p carries a weight, Wp(0 = Wp = 1), that reflects the
goodness of the position. Also, we know the number of games Gp where this
position has occured before. Finally, we have the average game result Rp
from these games for the player to move. (For example, Rp is 0.75 if we have
one won game and one drawn game for this position).

Even more deatiled is how the book-learning functions are applied once that
move has been played in-game, by the player.

Not to mention, the user selectable weighting of frequency, game-results,
and book-learning, as to how the engine will respond to the next occurance
of said position. Opening book move selection is a complex process, more
refined than mere frequency of appearance, and hopefully more successful in
ultimate result.



  #3   Report Post  
Old January 6th 04, 07:18 PM
Mike Ogush
 
Posts: n/a
Default Skewing results by use of one set of GM games.

On 5 Jan 2004 21:12:37 -0800,
(Dr. David Kirkby)
wrote:

I've been thinking about something and would appreciate the comments
of others.

Most chess programs use an opening book, which one will typically
based on a set of GM games. My understanding is that the chess
analysis program whilst 'in book' will evaluate a position based on
the frequency of occurance of the moves from this list of GM games,
since you have used that to create your opening book.

So if the opening move played by most grandmaster games is d4, then
the chess program will give the highest score if white plays d4,
rather than says a4, which would be a rare opening move for a
grandmaster. If the most common reply by black is Nf6, and one plays
that in a game, the chess engine will think that is black's best
response. Is my understanding of this correct?


Many chess programs also factor in the performance of a move when
selecting. Also some programs (Fritz for one) also have the ability
to "learn" from the games they play and update the opening book
accordingly.

If one then uses a set of GM games as a 'reference database' in a
database program (chessbase, scid etc) to evaluate the best opening
move, then if the same set of GM games is used, it will of course have
d4 as the most common opening move. Then when you evaluate the
position with a chess engine, it will give the highest score to white
if he plays d4. Likewise the database will show Nf6 as blacks most
common (and so one might infer best) response.

This suggests to me that the results of computer analysis of positions
get skewed if the ***same*** set of GM games are used both to create
an opening book in a chess analysis program AND as a reference
database in scid, chessbase or whatever.

Clearly with the first few opening moves it makes no difference if the
list of GM games is reasonably long. But as you progress further into
the game, the skewing of results might become significant.

Statistics never was my strongest subject, so I'd hate to guess when
this becomes significant, but I can't help feeling one develops a
false sense of what is the best move if both the chess engine and the
database program use the same set of GM games.

I'm using a set of about 27,000 GM games downloaded from the crafy ftp
site to create an opening book and also the same set in scid as a
reference database. I can't help feeling that this is not very
sensible. It seemed reasonably logical at first to use a set of high
quality games for both the opening book and the reference database,
but now I'm not so sure.

Comments ??

Dr. David Kirkby.


Yes, there is some skewing of a program's opening book based on the
games or game fragments being used to compile the book. In some cases
you may want the skewing:
* You are studying a particular opening and want to play a number of
practice games in that opening. You can do this by setting up a
position and starting the games from that position; but this can be
complex if you are studing a set of positions/variations (suchas as
the Sicilian Dragon) and still want the program to remain in book.
It is often easier to define the opening book so the program will
always stay within main book lines when using that book. [Aside:
Steve Lopez wrote a series of articles on using chess software to aid
in the learning of a new opening as part of his technical notes
series.]
* You are studying the repertoire of a particular player (a possible
future opponent) and you want the computer to only play the lines that
player plays regularly.

Also many people like to play lines that have some opening theory, but
are don't played much by GMs (the Morra and Blackmar-Diemer gambits
come to mind). Limiting the opening books to just games that GM plays
means that the program will be out of book much earlier than published
opening theory.

That being said, if you do want just GM games there are a number
downloadable collections of games where both players have a minimum
rating; I have seen this for 2500 and 2600 Elo at
www.uni-klu.ac.at/~gossimit/c/curious.htm.
Alternatively you could download a large collection (such as ChessLib)
and use a filtering program like pgn-extract to create your own
collection of games with a minimum rating of your own choosing. Hint:
If you do try the filtering approach be sure not to filter out games
of players that show no rating but are known to be of GM strength
(Capablanca, Alekhine, etc.)

Mike Ogush


  #4   Report Post  
Old January 7th 04, 01:18 AM
Dr. David Kirkby
 
Posts: n/a
Default Skewing results by use of one set of GM games.

"Derek Wildstar" wrote in message news:[email protected]_s51...
"Dr. David Kirkby" m wrote
in message m...

If the most common reply by black is Nf6, and one plays
that in a game, the chess engine will think that is black's best
response. Is my understanding of this correct?


Not exactly, Frequency of appearance is just one criteria of move
applicability in book.
Here is an example, from Rookie 2.0, of how opening moves are weighted:

"If n moves are available to be played then the probability for a move m
(leading to p) to get chosen is proportional to P(m) = (1-Wp) x Gp x
(1-Rp). This way we favor moves with higher winning expectations and also we
prefer to choose moves that will keep us in the book as long as possible."


snip

I accept from your description that the chess engine will not play
moves in a way directly proportial to their occurance in the games in
its database. You have clearly shown that, although your state one
factor will be the number of wins achieved by playing a particular
move.

But will the chess engine evaluate the position of the board based on
the data in that database, if the position is still in its book ?? I
assume there is no random number generator called score the position.

If so, then that still leaves the possibility that you errornously
believe the best move to play from Position A is move B, basing this
on

a) The chess engine thinks its the best move AND
b) it lead to the most wins by GMs from that position.

Based on (a) and (b), one might conclude incorrectly move B is best,
simply because both the chess engine's opening book and your database
are both based on an identical set of games.

Again, I could be completly off track here. It is certainly not
something I know much about (as I've clearly shown), but the issue did
get me thinking a bit.

Dr. David Kirkby
  #5   Report Post  
Old January 7th 04, 05:47 AM
Derek Wildstar
 
Posts: n/a
Default Skewing results by use of one set of GM games.


"Dr. David Kirkby" m wrote
in

I accept from your description that the chess engine will not play
moves in a way directly proportial to their occurance in the games in
its database. You have clearly shown that, although your state one
factor will be the number of wins achieved by playing a particular
move.


So far so good, stipulating that frequency is just one factor in the
decision making process about the 'goodness' of the move in book.

But will the chess engine evaluate the position of the board based on
the data in that database, if the position is still in its book ?? I
assume there is no random number generator called score the position.


Based on the prior written description of Rookie 2.0's methodology in-book,
yes there is a form of evaluation of the position's 'goodness', which is not
the exact same evaluation of the positition out-of-book. For this evaluation
is again just a single factor, weighted, in determining the probability of a
book-move being played... i.e. frequency of occurance.

I suspect that in cases where a 50/50 result occurs, some randomness is to
be expected, of course, this position when seen again, will no longer result
in that same 50/50 evaluation, if the result of that last occurance is now
weighted as well.


If so, then that still leaves the possibility that you errornously
believe the best move to play from Position A is move B, basing this
on

a) The chess engine thinks its the best move AND
b) it lead to the most wins by GMs from that position.

Based on (a) and (b), one might conclude incorrectly move B is best,
simply because both the chess engine's opening book and your database
are both based on an identical set of games.

Again, I could be completly off track here. It is certainly not
something I know much about (as I've clearly shown), but the issue did
get me thinking a bit.

Dr. David Kirkby


I think you have unrealistic expectations of opening books. And before you
think I'm the last word on books, I'm certainly not, but hopefully any
errors will be corrected before they give you any bad-habits!

Are you assuming that a decided upon book-move is the best move in that
position? This is the theoretical aim of pattern-recognition, but the
reality falls from the mark. Especially when you are under the impression
that a population of GM games is the appropriate foundation for a book.
Perhaps basing a book on a set of games is correct for opening practice, or
specialized training, but as a general purpose aid to increasing the overall
strength of a program, it's not correct.

A general purpose opening book instead should have a hand picked set of
positions that covers most, if not all, tournament variations, most, if not
all, principle variations and Main Lines, and known traps, pitfalls and
blunders, especially for lines that can cause the depth of search to stay
depressingly shallow, keeping the correct line hidden.

None of those conditions are met by including a list of games, however
broad, especially by high rated players. They simply do not go into lines
that amateurs do. I think you know all this already, or are close to knowing
this, and you are wondering what exactly that DB of 27K games is going to do
for you...

I do not think incorporating that list into a general purpose book will do
you much good. However, not willing to discount it as a worthwhile
experiment, may I offer a suggestion, to put the issue to a test:
Incorporate the games into an existing book, and then run an engine-match
tournament with four players:

1) General Purpose Book
2) General Purpose Book + GM Games
3) GM Games
4) No Book

See where I'm going with this...?







  #6   Report Post  
Old January 8th 04, 06:29 PM
Mike Ogush
 
Posts: n/a
Default Skewing results by use of one set of GM games.

On Wed, 07 Jan 2004 05:47:06 GMT, "Derek Wildstar"
wrote:


"Dr. David Kirkby" m wrote
in

I accept from your description that the chess engine will not play
moves in a way directly proportial to their occurance in the games in
its database. You have clearly shown that, although your state one
factor will be the number of wins achieved by playing a particular
move.


So far so good, stipulating that frequency is just one factor in the
decision making process about the 'goodness' of the move in book.

But will the chess engine evaluate the position of the board based on
the data in that database, if the position is still in its book ?? I
assume there is no random number generator called score the position.


Based on the prior written description of Rookie 2.0's methodology in-book,
yes there is a form of evaluation of the position's 'goodness', which is not
the exact same evaluation of the positition out-of-book. For this evaluation
is again just a single factor, weighted, in determining the probability of a
book-move being played... i.e. frequency of occurance.

I suspect that in cases where a 50/50 result occurs, some randomness is to
be expected, of course, this position when seen again, will no longer result
in that same 50/50 evaluation, if the result of that last occurance is now
weighted as well.


If so, then that still leaves the possibility that you errornously
believe the best move to play from Position A is move B, basing this
on

a) The chess engine thinks its the best move AND
b) it lead to the most wins by GMs from that position.

Based on (a) and (b), one might conclude incorrectly move B is best,
simply because both the chess engine's opening book and your database
are both based on an identical set of games.

Again, I could be completly off track here. It is certainly not
something I know much about (as I've clearly shown), but the issue did
get me thinking a bit.

Dr. David Kirkby


I think you have unrealistic expectations of opening books. And before you
think I'm the last word on books, I'm certainly not, but hopefully any
errors will be corrected before they give you any bad-habits!

Are you assuming that a decided upon book-move is the best move in that
position? This is the theoretical aim of pattern-recognition, but the
reality falls from the mark. Especially when you are under the impression
that a population of GM games is the appropriate foundation for a book.
Perhaps basing a book on a set of games is correct for opening practice, or
specialized training, but as a general purpose aid to increasing the overall
strength of a program, it's not correct.

A general purpose opening book instead should have a hand picked set of
positions that covers most, if not all, tournament variations, most, if not
all, principle variations and Main Lines, and known traps, pitfalls and
blunders, especially for lines that can cause the depth of search to stay
depressingly shallow, keeping the correct line hidden.

None of those conditions are met by including a list of games, however
broad, especially by high rated players. They simply do not go into lines
that amateurs do. I think you know all this already, or are close to knowing
this, and you are wondering what exactly that DB of 27K games is going to do
for you...

I do not think incorporating that list into a general purpose book will do
you much good. However, not willing to discount it as a worthwhile
experiment, may I offer a suggestion, to put the issue to a test:
Incorporate the games into an existing book, and then run an engine-match
tournament with four players:

1) General Purpose Book
2) General Purpose Book + GM Games
3) GM Games
4) No Book

See where I'm going with this...?


I wanted to say a bit more about the problems of using a game list to
try to generate a perfect opening book:

1. As Steve Lopez (in the Chessbase Technical Notes) has pointed out
statstics (regarding the frequency a particular move has been played
and that moves performance) can be deceptive.

Say in some opening line White has played 14.Nxf6 in 20 games with a
60% performance rating and as a result Black may have even abondoned
the line that led to the position where white plays 14.Nxf6, then some
enterprising GM deeply analyzes Blacks subsequent play and finds a way
for Black to get an advantage by force. Even if he wins the game
White's performance will only drop to 57% - falsely indicating that
14.Nxf6 is still a pretty good move.

2. One way of getting around the statistical problem is to use a
feature of Bookup called backsolving. Essentially backsolving starts
from the end of all games in a collection and propagates the
evaluation back up the move tree (using +- as the evaluation when
white wins, and -+ as the evaluation when black wins). So if you
backsolved the 21 games with 14.Nxf6 the game where the GM forced an
advantage to Black and won the game would override other games and
give 14.Nxf6 an evaluationj of -+.

Unfortunately backsolving has certain limitations when it relies
solely on game results: A player can lose on time in a won position.
A player could play a three-fold repition allowing a draw in a won
position. A player can reach a won position and through a series of
mistakes throw away the win and even the draw and eventually lose.
This can be accounted for by doing extensive analysis of games rather
than relying on their results to get a good evaluation for a given
position.

3. Lastly even if the evaluations for a moves are completely accurate,
they can result in situation where the computer as white has an
advantage (+/-) or even a won game (+-) when it leaves its opening
book, but the program does not know how to play the subsequent
positions. This is often true if the advantage is positional and
requires a player to do long term strategic planning in order to
realize the win; this is eactly what most programs are notoriously
bad at. I have seen humourous situations where computers drop out
book with a position that a human being recognizes as completely won
and the computer's evaluation of the position is that its opponent has
a winning advantage.

Because of the above, many commercial chess program vendors have an IM
or even a GM on staff to help tweak the opening book. A few years ago
the chess site for Rebel would include some notes about the tweaking
done to the opening book for Rebel; I don't know if they still do
this. In some of the man-machine matches of the last few years with
Kaspraov and other GMs the reports on the matches would talk about
how the opening book needed to be modified to stay away from close
positions where a GMs ability at strategic maneuvering would give him
the advantage.

When a computer is playing humans below master level the errors in the
opening book aren't so critical. Although, we do see people bragging
how they can beat a strong program time after time, which is only
because they have found a new flaw in the programs opening book.



Mike Ogush
  #7   Report Post  
Old January 8th 04, 08:09 PM
Derek Wildstar
 
Posts: n/a
Default Skewing results by use of one set of GM games.


"Mike Ogush" wrote in message
...

relevant snippage

Mike Ogush


Thanks for those points Mike, I hope the OP appreciates the info. I know I
do. I've taken my own advice and using the Chesspartner (5.3) interface,
I've assembled a few personalities based on their bonus book downloads
(Thank you Lokasoft), so far I have:

Classic Openings
Gambits
Grand Master Games
LCHESS
Modern Openings
John Nunn's Picks (!)
Sharp Openings

I'll set up CP's "Engine Research Tool" and using Ruffian, sic them upon
each other. I do not have any expectations of opening revelation, but I'm
curious to see if things play out as expected: Gambit losing to GM, Classic
losing to Nunn...

Even better, see which book(s) play best against people, and which play best
against AI. I suspect the Modern Book and Ruffian will have the most
difficult time together, Ruff has deep affinity for the principle variation,
and freaky indirect openings frighten it into shallowness.





  #8   Report Post  
Old January 9th 04, 06:14 PM
[email protected]
 
Posts: n/a
Default Skewing results by use of one set of GM games.

(Mike Ogush) wrote in message ...
On 5 Jan 2004 21:12:37 -0800,
(Dr. David Kirkby)
wrote:

I've been thinking about something and would appreciate the comments
of others.

Most chess programs use an opening book, which one will typically
based on a set of GM games. My understanding is that the chess
analysis program whilst 'in book' will evaluate a position based on
the frequency of occurance of the moves from this list of GM games,
since you have used that to create your opening book.

So if the opening move played by most grandmaster games is d4, then
the chess program will give the highest score if white plays d4,
rather than says a4, which would be a rare opening move for a
grandmaster. If the most common reply by black is Nf6, and one plays
that in a game, the chess engine will think that is black's best
response. Is my understanding of this correct?


openings story settings Video Camera Samsung Dealers Norway multi
picking cracking spying warez games espionage bugs counter measures
cable tv test chips red box computer hacking cable tv converter boxes
videocipher smart card chip computer credit cards phrozen crew cracks
telephone decoder descrambler wireless ecm programming telecode scanner
crypto digital communication phone phreaking red boxes cable tv hacking
smart card hacking surveillance equipment my-deja.com red my nu zite box
machines troubleshoots ccctournament header key fields "Pea soup" green
t616 rate my camel toe Keehn chessbase converter radi caraib Techman Head
clothing jewish nobel prize winners beaulieu alaska landscaping fmccoy
board +british ccct island custom
tripp plane sore
casio mr g
watches snore the hispanics i river mp3
debt
http://amateurschach.de negotiation british pop charts terminator
stocks theatre in the uk landlord representation @my-deja.com florida
evasion CCCT refurbished computers phrocrew magnetostriction amplify
himage dvd 8800 sulzer orthopedics attorneys florida www.ancestry.com
microsystems cellularseparation anxiety "business
decreased libido refurbished cdr downloads st-europe @my-deja.com currants
worksheet organization pdf puddle clay pile
Compose
colonic irrigation uk jordan La Maison Picassiette supply small looks
almaty "villa monaco" overland hairstyles
design
weld glasses 200246 2500 ebay nikon cf-d100 arizona income property abc
player christmas cat florist kent panasonic portable player locksmith
motivation letter winDVD restart.com unpot emortenson Hammargren article
papers rpg computer games chess looney tunes back lunatic alcohol
hadieh tehrany ladiest emmanuel avi test blade computer telephone with
debt consolodation flight of the navigator mobile ringtone fntbl skipper
lyrics hard knocked life apartheid cure
hampton
keywords unops Philippine Western Visayas mediacleaner new generation
water rats series Karl Mas Sveiby qnx cannot finalize dvd location hitler

Many chess programs also factor in the performance of a move when
selecting. Also some programs (Fritz for one) also have the ability
to "learn" from the games they play and update the opening book
accordingly.

If one then uses a set of GM games as a 'reference database' in a
database program (chessbase, scid etc) to evaluate the best opening
move, then if the same set of GM games is used, it will of course have
d4 as the most common opening move. Then when you evaluate the
position with a chess engine, it will give the highest score to white
if he plays d4. Likewise the database will show Nf6 as blacks most
common (and so one might infer best) response.

This suggests to me that the results of computer analysis of positions
get skewed if the ***same*** set of GM games are used both to create
an opening book in a chess analysis program AND as a reference
database in scid, chessbase or whatever.

Clearly with the first few opening moves it makes no difference if the
list of GM games is reasonably long. But as you progress further into
the game, the skewing of results might become significant.

Statistics never was my strongest subject, so I'd hate to guess when
this becomes significant, but I can't help feeling one develops a
false sense of what is the best move if both the chess engine and the
database program use the same set of GM games.

I'm using a set of about 27,000 GM games downloaded from the crafy ftp
site to create an opening book and also the same set in scid as a
reference database. I can't help feeling that this is not very
sensible. It seemed reasonably logical at first to use a set of high
quality games for both the opening book and the reference database,
but now I'm not so sure.

Comments ??

Dr. David Kirkby.


Yes, there is some skewing of a program's opening book based on the
games or game fragments being used to compile the book. In some cases
you may want the skewing:
* You are studying a particular opening and want to play a number of
practice games in that opening. You can do this by setting up a
position and starting the games from that position; but this can be
complex if you are studing a set of positions/variations (suchas as
the Sicilian Dragon) and still want the program to remain in book.
It is often easier to define the opening book so the program will
always stay within main book lines when using that book. [Aside:
Steve Lopez wrote a series of articles on using chess software to aid
in the learning of a new opening as part of his technical notes
series.]
* You are studying the repertoire of a particular player (a possible
future opponent) and you want the computer to only play the lines that
player plays regularly.

Also many people like to play lines that have some opening theory, but
are don't played much by GMs (the Morra and Blackmar-Diemer gambits
come to mind). Limiting the opening books to just games that GM plays
means that the program will be out of book much earlier than published
opening theory.

That being said, if you do want just GM games there are a number
downloadable collections of games where both players have a minimum
rating; I have seen this for 2500 and 2600 Elo at
www.uni-klu.ac.at/~gossimit/c/curious.htm.
Alternatively you could download a large collection (such as ChessLib)
and use a filtering program like pgn-extract to create your own
collection of games with a minimum rating of your own choosing. Hint:

If you do try the filtering approach be sure not to filter out games

of players that show no rating but are known to be of GM strength

(Capablanca, Alekhine, etc.)

Mike Ogush

  #9   Report Post  
Old January 10th 04, 12:31 AM
Simon Waters
 
Posts: n/a
Default Skewing results by use of one set of GM games.

Dr. David Kirkby wrote:

This suggests to me that the results of computer analysis of positions
get skewed if the ***same*** set of GM games are used both to create
an opening book in a chess analysis program AND as a reference
database in scid, chessbase or whatever.


Most of the programs I've seen may choose opening moves based on
frequency, but they don't typically use the book information in the
search (other than perhaps to order move - which would be a subtle form
of influence indeed).

So the analysis the computer returns is probably tactically sound unless
the database person has tried to get clever.

Currently GNU Chess has a fairly crude method for using it's book games,
which picks the best scoring moves, with a popularity cut off (b4 has an
excellent score but hasn't been played enough). This very crude
algorithmn occaisonally throws a wobbly, but against myself it often
leaves it with a winning advantage by the time it leaves booksigh.

Whilst GNU Chess now thinks (in it's opponents time only) in the
openings - this is purely to prevent it from reaching a middle game with
no analysis - which cost it a blitz game against a FIDE master where it
made a really bad blunder due to having such a complex position it was
unable to complete even a preliminary search of the position in time.

Contary to "folk wisdom" computers often do fairly well without an
opening book, and I'm tempted to go the complete opposite and just use
opening books as a source of move ordering information for the search
algorithmn. But I think the time penalty against players with carefully
prepared books may be too large.

-----BEGIN PGP SIGNATURE-----
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iD8DBQE//0fpGFXfHI9FVgYRAnPiAJ4kyk4vYVRzfvZhSZJdrwhDF1/92ACgiGjP
UMGZZh9u0uCX0TRhm1tZJ6c=
=CO/T
-----END PGP SIGNATURE-----

  #10   Report Post  
Old January 10th 04, 12:19 PM
Dr. David Kirkby
 
Posts: n/a
Default Skewing results by use of one set of GM games.

(Mike Ogush) wrote in message ...

I wanted to say a bit more about the problems of using a game list to
try to generate a perfect opening book:

1. As Steve Lopez (in the Chessbase Technical Notes) has pointed out
statstics (regarding the frequency a particular move has been played
and that moves performance) can be deceptive.

Say in some opening line White has played 14.Nxf6 in 20 games with a
60% performance rating and as a result Black may have even abondoned
the line that led to the position where white plays 14.Nxf6, then some
enterprising GM deeply analyzes Blacks subsequent play and finds a way
for Black to get an advantage by force. Even if he wins the game
White's performance will only drop to 57% - falsely indicating that
14.Nxf6 is still a pretty good move.


Yes I can see that.

2. One way of getting around the statistical problem is to use a
feature of Bookup called backsolving. Essentially backsolving starts
from the end of all games in a collection and propagates the
evaluation back up the move tree (using +- as the evaluation when
white wins, and -+ as the evaluation when black wins). So if you
backsolved the 21 games with 14.Nxf6 the game where the GM forced an
advantage to Black and won the game would override other games and
give 14.Nxf6 an evaluationj of -+.

Unfortunately backsolving has certain limitations when it relies
solely on game results: A player can lose on time in a won position.
A player could play a three-fold repition allowing a draw in a won
position. A player can reach a won position and through a series of
mistakes throw away the win and even the draw and eventually lose.
This can be accounted for by doing extensive analysis of games rather
than relying on their results to get a good evaluation for a given
position.

3. Lastly even if the evaluations for a moves are completely accurate,
they can result in situation where the computer as white has an
advantage (+/-) or even a won game (+-) when it leaves its opening
book, but the program does not know how to play the subsequent
positions. This is often true if the advantage is positional and
requires a player to do long term strategic planning in order to
realize the win; this is eactly what most programs are notoriously
bad at. I have seen humourous situations where computers drop out
book with a position that a human being recognizes as completely won
and the computer's evaluation of the position is that its opponent has
a winning advantage.

Because of the above, many commercial chess program vendors have an IM
or even a GM on staff to help tweak the opening book. A few years ago
the chess site for Rebel would include some notes about the tweaking
done to the opening book for Rebel; I don't know if they still do
this. In some of the man-machine matches of the last few years with
Kaspraov and other GMs the reports on the matches would talk about
how the opening book needed to be modified to stay away from close
positions where a GMs ability at strategic maneuvering would give him
the advantage.


When a computer is playing humans below master level the errors in the
opening book aren't so critical. Although, we do see people bragging
how they can beat a strong program time after time, which is only
because they have found a new flaw in the programs opening book.



Mike Ogush


Thanks for the detailed explanation Mike. It is clear there is a lot
more to the opening book than I thought. From my own point of view,
crafty can always beat me unless I make serious attempts to weaken it,
so making a strong machine to play aginst does not bother me that
much. I was however keen to use the database programs and chess engine
in combination to find good moves to play in the opening. But its
clear my knowledge it too limited at this point to really comment
further on any skewing that may or may not occur.

David Kirkby
Reply
Thread Tools
Display Modes

Posting Rules

Smilies are On
[IMG] code is On
HTML code is Off
Trackbacks are On
Pingbacks are On
Refbacks are On


Similar Threads
Thread Thread Starter Forum Replies Last Post
Rauzer Attack Rik rec.games.chess.analysis (Chess Analysis) 3 July 14th 04 09:28 AM
Best Ever Chess Games (Soltis list?) Gregory Topov rec.games.chess.analysis (Chess Analysis) 0 January 20th 04 03:00 PM
Jeff Sonas about computers surpassing humans Sterten rec.games.chess.computer (Computer Chess) 15 October 16th 03 06:03 PM
English Opening Sam Stark rec.games.chess.analysis (Chess Analysis) 3 July 17th 03 09:49 PM
English Opening Sam Stark rec.games.chess.computer (Computer Chess) 1 July 17th 03 02:56 AM


All times are GMT +1. The time now is 02:25 PM.

Powered by vBulletin® Copyright ©2000 - 2019, Jelsoft Enterprises Ltd.
Copyright 2004-2019 ChessBanter.
The comments are property of their posters.
 

About Us

"It's about Chess"

 

Copyright © 2017