Author: David H. McClain
Date: 11:47:07 03/19/05
Go up one level in this thread
On March 18, 2005 at 10:34:56, Norm Pollock wrote: >On March 18, 2005 at 09:07:50, Dimitris Poulos wrote: > >>I wouldn't think so anymore. Since almost any game contains blunders and Norm's >>discussion on percentages of won games some weeks ago, I would consider as fair >>a book that keeps this percentages. Or a book with all GM games except the >>innovations better. >>Just to ponder. >> >>Dimitris > >All games contain blunders. No person or computer knows how to play the perfect >game. Even drawn games are not perfect. Assuming these three premises, how do >you make up the best possible opening book? > >First get a large collection (100,000+) of long time control games between >(human) professional players (elo 2400+ perhaps). Second, use recent games, >especially games from the "computer age" (say year 2000+). Professional players >today practice their openings on computer and look for innovations from the >computer. So by using recent games, you are getting the advantage of computer >analysis to some extent. Third, eliminate shorter games (20 moves or less >perhaps) that probably have an opening blunder or were grandmaster draws. >Fourth, check the collection of games to be sure that all games have results, >and that all games have elo ratings within the range you want. > >An often used approach is to filter out "bad" moves by indicating each >individual "bad" move with some sort of NAG. But to do this right, you would >need to be sure which moves are "bad". And you would have to go through 100,000 >games, move by move. When the engine sees the nag, it will avoid that move. > >The approach I recommend is to use separate books for white and black. Then use >"quantity" of repitition to effectively discard questionable moves. If you have >a book of say 60,000 games involving professional players, then choosing only >those moves that have occurred 12 times (0.02%) or more seems reasonable. >Unfortunately most engines and interfaces do not use separate books so it is >even more important in these cases to use a high repitition rate (at least >0.02%). > >How do you make separate books? For white, take only games white has won. There >still could be blunders in the openings that black did not see, but that will be >minimal since white won the game against a professional opponent. > >For black, there will not be as many games that black won compared to white. So >use games that black either won or drew. Similarly there will be blunders that >white did not see, but they will be minimal because black did not lose. > >If you can put effective nags into the book(s) as well, then it is even better. > >Btw, if you go this route, you might want to look at a collection of games >(normbk02.pgn) and a pgn utility (elopgn.exe) that I have uploaded at: > >http://www.crafty-chess.com/down/Pollock/ Norm, And how do you choose the corresponding color book when you are playing "autocolor" machine games? DHM
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.