Author: Norm Pollock
Date: 07:50:01 10/14/04
Engines say they have "book learning". But is all "book learning" the same? Is there a "right" way to implement "book learning"? The way I understand book learning is that it keeps a statistical record or wins/draws/losses for each opening, and adds this to the wins/draws/losses that came in the original opening book. I have two opinions on the use of book learning. (1) Do not use book learning if the book is shared. Suppose engine A plays engine B, both sharing the same opening book. In the aggregate, the book learning will be a push: either 1 win + 1 loss, or 2 draws. Nothing is learned in the aggregate. However if the book learning is split by engine v engine, then the learning would tell A to play the opening again against B, and B should avoid that opening against A in the future. (2) Book learning has to be split into book learning against each different opponent. Engine A may be able to consistantly beat engine B with a specific opening, but engine A may consistantly lose to engine C with the same opening. What has to be learned is for engine A to use that opening against engine B, but not against engine C. If the learning is not split by opponent, the wins over B will be balanced by the losses to C, and nothing would be learned in the aggregate. Now when an engines claims to have "book learning", I don't know if the book learning is by aggregate or is split by engine v each opponent. I wish that this information is made available by each author so that I can make the right decision as to whether or not to use the "book learning".
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.