Author: Jon Dart
Date: 20:56:22 01/20/03
Go up one level in this thread
This was discussed a little bit on channel 64 during the tournament. Most amateurs are using a large PGN collection as book data, often with a smaller hand-tuned book that is probed first, and that can be used to steer the program into favorable lines or avoid unfavorable ones. This is often an adequate approach, but I gotten to mistrust the use of more or less raw PGN data as input. If you use game results as a guide to how good an opening move is, you can be mislead, because sometimes (for example) a game winds up lost when the losing side really had a strong advantage out of the opening. Also move frequency is misleading. You can see 10 games in a game collection that have the same move in a particular position. But the 11th game demonstrates that there's a stronger move. A book tuned to play more frequently occurring moves misses this. There are ways to reduce these problems, but basically if you feeding bad data into the book creation process, you will get bad opening moves in your book - they won't all be bad, but some will be. Arasan now ships with only the hand-tuned part of the book. You can download a bigger book that is built out of game collections automatically, but I don't generally use the big book myself. I have weeded bad lines out of the hand-tuned book by watching it play online, running long-duration auto-matches offline against various programs, and by some opening research in books and game databases. So it makes fewer gross mistakes now, and I've also had it avoid some openings on general principles, because it just doesn't seem to play them well. I happen to enjoy this activity, so I spend quite a bit of time on it. It's more effort than most people will do, but it falls quite a bit short of what professional book creators do. A major disadvantage of my book is that it is not very large: about 33,000 moves at present. Having a small book is especially a disadvantage if you have GMs pounding on your program all day, like Crafty does, because they have a lot of opening knowledge already, and learn pretty fast where the holes are. The really good opening books are simultaneously large, carefully tuned, and checked for errors. But the effort required to build that sort of book is very substantial, and it needs continuous updating if it is to continue to be useful against high-level competition. --Jon
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.