Author: Robert Hyatt
Date: 09:03:29 05/20/02
Go up one level in this thread
On May 19, 2002 at 18:00:02, Roy Brunjes wrote: > >I wonder how book learning is implemented in various commercial chess programs >on the market. Has this been discussed before? If so, can someone point me to >a link here at CCC about this topic? > >As an example: Program X plays the Ruy Lopez (as White) vs an opponent. Let's >say that the game was played at Game in 30 minutes. The program ultimately >loses the game after 75 moves. With book learning, will it refuse to open 1. e4 >ever again (assuming the program is set to never repeat an opening line it loses >with)? Does the time control of that game get noted in the book learning >process? If a program is unable to win with a line at game in 30 minutes, who >knows if it might do better at Game in 5 minutes with the same line, or game in >2 hours instead? If you are an ICCA member, you can find a copy of a paper I wrote on book learning there. What I do is pretty similar to what anybody has to do if they implement book learning. The "weights" might be different, but the overall approach is pretty well-defined and straightforward. If you aren't an ICCA member, I think I have an electronic (ASCII) version I could send. > >With a defeat on its record with 1. e4 ... how does it know that the source of >its defeat was, perhaps 2. Nf3 (after 1 ... e5 by Black). Or does it assume >that the 22nd move (I'm pulling that number out of the air for this example), >which is the last book move it has, is the offending move and the program will >willingly repeat that line up to move 22 and then vary (either select another >book move if there is one, or if there is not another book move at that point, >just start thinking in hopes of finding a better move on its own)? > >And, if that deviation from book (i.e. avoiding the known "bad" 22nd move in its >book) results in another defeat, does the program then mark the 21st move in its >book as suspect -- and so on ... working its way back up toward 1. e4 only after >a huge number of games is played (all of which in my contrived example) lead to >losses for the program? > >Also, do programs learn differently if they are White with a given opening line >vs if they are Black in the same opening line? > >Yes, I realize that there are probably many different ways to implement book >learning, but hopefully my example if generic enough that someone can describe a >plausible way to handle it, even if no specific programmer for any single >commercial product responds to my questions. > >Thanks for your thoughts, > >Roy
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.