Author: Norm Pollock
Date: 09:08:47 01/28/05
Go up one level in this thread
On January 28, 2005 at 11:45:29, Uri Blass wrote: >On January 28, 2005 at 10:55:23, Norm Pollock wrote: > >>On January 28, 2005 at 10:44:59, Matthias Gemuh wrote: >> >>>On January 28, 2005 at 09:19:47, Norm Pollock wrote: >>> >>>>Stefan, >>>> >>>>I would like to ask you why the commercial engines do not use a double book - >>>>one book for moves as white and another for moves as black. By using one book, >>>>based a single collection of games, black for example, is basing its opening on >>>>games that black lost. >>>> >>>>-Norm >>> >>> >>>The book is basically 2 books in file because the side to move is part of the >>>hash key used to probe the file for moves. >>> >>>/Matthias. >> >>From an operational standpoint I agree that the book is really 2 books. >> >>I am considering how the book (or 2 books from an operational standpoint) is >>created. It is created from pgns that have games of 1-0, =-= and 0-1 in them. So >>white is using "its" book based on pgns that include 0-1, and black is using >>"its" book based on pgns that include 1-0. >> >>Uri's comment about Fritz shows that Fritz handles this by commenting out some >>weak moves. But how complete and accurate could this be? Isn't it better to >>first base an opening book on games white (for example) won or drew against a >>highly rated opponent, and then require a high frequency for any move to be >>entered into the book? > >This is almost the way that I do it in movei > >The public book for white is based only on games that white won and I need >minimal frequency of 5 times. >the public book for black is based only on games that black got at least a draw >and I need minimal frequency of 5 times. > >I think that the games are all games of GM's that I downloaded from the >internet. > >I do not think that the way that I do it is optimal and I doubt if using >statistics to decide about moves is a good idea because it is possible that a >move was considered as good in the past and later some refutation was found. > >I think that some expert who write the book manually is better and I think that >sandro writes the commercial book for shredder manually so I expect shredder uci >to have better book than the book that movei has. > >Uri I have a few questions about how you made your book. Have you tested your engine with book vs your engine without book, preferably at short & blitz time controls where a book has extra importance? Does your book really make a measurable improvement? Also did you test your engine with/without book against a gauntlet of 4 engines of comparable strength using their books? What is the minimum elo of games in your database for your book? How many games are in the database? What are the dates on those games? Would you consider enlarging the database you based your book upon? Would you consider a higher frequency of occurrences (like around 12)? Did you check out the database I put up on Peter Skinner's crafty-chess.com ? It has about 60,000 games, elos 2400+, no duplicates, long-time controls, min 20 moves, and relatively recent (since year 2000).
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.