Author: Eric Somao
Date: 22:50:16 03/13/04
Go up one level in this thread
Thanks a lot for ur advices... and ur explanations. It s very precious. Regards Eric On March 14, 2004 at 00:19:22, Mike S. wrote: >On March 13, 2004 at 10:26:01, Eric Somao wrote: > >>I' ve collected more than 15 000 games played by Shredder 7 and 8 in more than >>20 minutes/game and against the 4 best current engines (Fritz, Junior, Hiarcs >>and Ruffian). I look for the best way of building the best possible book for >>Shredder. >> >>So I have few questions: >> >>- Must I use only these games or do I have to use a small book (remis.ctg from >>Mr Utzinger) which would be a core in which I would import my base? > >I wouldn't mix that data with another book, because if you'd do that, the >book(s) would sometimes - or often - play continuations which have nothing to to >with the Shredder games. - Of course it's a problem when opponents play >something else than the programs did in your collection. But in 15.000 games, at >least most "mainstream" openings will be covered. Also, if these have most often >been played with "optimal" books settings resulting in a narrow repertoire so to >speak, you can be confident that the opponents for your new book will again have >chosen optimal (or you could do that of course, if you control both progs). > >>- Secondly: after the importation, do I have to make the book learn from the >>base or is it useless? > >If you build a new book tree based on these games, it's of course not necessary >to learn from the same games then, afterwards. All the necessary statistical >data will already be in the book, from the generation process. > >I suggest to generate **two** books, one for Shredder with white based on >Shredder's white wins, and for black resprectively. > >Because if you merge both wins and losses by Shredder into the same book, it >will simply prefer variants for White or Black, and these won't be based on >Shredder only but on the oppenents as well. In other words, if for example >Junior was successfull with a specific variant often, the book would prefer it >just because the statistics are good, even if it may not fit Shredder. This >wouldn't be engine specific. But you can achieve that by dividing the database >into two: White wins, and Black wins by Shredder (I suggest to leave the drawn >games to both databases). > >The problem in practise is, that you cannot switch books automatically within an >engine tournament or during automatical play on servers. An engine match can be >played in two parts though, without alternating colors each. > >If you need to have all in *one* book only, I see no easy way to determine which >moves were succsessfull for Shredder (more), and which more for the opponents... > >Regards, >Mike Scheidl
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.