Author: Robert Hyatt
Date: 10:18:29 10/31/00
Go up one level in this thread
On October 31, 2000 at 00:16:49, Gordon Rattray wrote: >How simple or complicated do chess programs implement opening book learning? Is >there much variation amongst the top programs? > >For example, when I use Fritz/Junior, I often pair them against opponents of >various strengths. For weak opponents they always win. By blinding learning >that their chosen opening was a success, are they really improving their future >chances against stronger opponents? Or is there more to it? > >I understand that it is not for the program to know straight off whether it is >playing a strong or weak opponent, but can't some factors from the remainder of >the game be used? Can't it get an idea of whether the opponent played badly? >And then take this into account? > >I've read here before that the initial evaluation out-of-book isn't vital. More >important is whether the *type* of position suits the program's style. So, from >this, I guess it's not the starting value but the evaluation trend that needs to >be examined to see if the opening was good or bad. Comments? > >Another question... If the opponent loses on time, but I continue anyway and >Fritz/Junior "loses on the board", is the opening book updated for a win or >loss? Here, I think it needs to be a loss regardless of the clock. I did have >a look at the opening book values and I think this is being treated as a win?! > >Gordon Fritz uses a slightly modified learning algorithm taken from Crafty. Which means the following is probably true for it, as the following is true for Crafty. 1. Crafty doesn't learn from "wins". Because they are less trustworthy. 2. Crafty does learn from losses and won't repeat a line that loses, ever. It might repeat it for the other side of course. Since it was good enough to beat itself... 3. Crafty learns "good" things by watching the eval when it leaves the book, and it "studies" this evaluation for the first 10 non-book moves, to try to determine whether the opening led to a + position or a - position. It also factors in the ratings for it and the opponent so that good scores against a weaker opponent don't influence the book as much as good scores against an even better opponent.
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.