Author: Jay Scott
Date: 08:12:33 08/21/98
Go up one level in this thread
On August 20, 1998 at 23:14:49, Steffen Jakob wrote: >On my page http://www.jakob.at/steffen/hossa.html I describe the book learning >approach which I use for my program "Hossa". Hossa uses a form of offline opening book learning. As far as I know, all other chess programs use online book learning, which is weaker in principle. Hooray! A pioneer! Othello programs have always used offline learning. For some reason, maybe because the first thing they tried worked OK, chess programmers have never picked up those techniques (as far as I know). Maybe it's a form of snobbery: of course mere othello techniques could never work in a real game, like chess. Your web page doesn't fully describe the learning method. Are you backing the actual game result up the tree? If so, Hossa's method is very similar to what the champion othello program Logistello did for years (Logistello is retired now). There is a risk of learning from bad games--of thinking that a position is bad because later on one side made a silly blunder and lost. Learning that "all moves lose" (say, from the starting position after playing against a stronger opponent) is an example of this. Logistello solved it by correcting all endgames with exact calculation, but that method isn't feasible for chess. :-) Logistello went further: it searched for winning deviations in lines that seemed to lose. In other words, it discovered opening novelties on its own. I'd love to see how well that works in chess! For Logistello info, see my page: http://forum.swarthmore.edu/~jay/learn-game/systems/logistello.html I think a fully satisfactory opening book learner would combine the two sources of information that a program has about how good a position is: (1) the evaluation of the position as returned by the search, and (2) the outcome of games played from that position. So far, all the learners I've seen use primarily one or the other; apparently nobody's figured out how to combine the two into one number in a principled way. Jay
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.