Author: Bruce Moreland
Date: 17:14:04 04/22/98
Go up one level in this thread
On April 22, 1998 at 18:46:08, Jan-Frode Myklebust wrote: >So then Deep Blue isn't that good at chess in general, only at beating >G. Kasparov under special conditions...? At least it can never become >world champion without either learning to avoid playing the same loosing >line again and again, or solve the game... Put it this way, if your problem domain is a six-game match, with a day or two between games, then you don't need machine learning, since you can modify the program by hand between games. If your problem domain is hundreds of games per day, consisting of a series of matches of between one and fifty games duration, against low-rated people who are quite happy to play the same line over and over again (losing one rating point each time) until they win (gaining 31 points from you at that moment), then repeating this won line over and over until you notice and pull the plug, then yes, machine learning is probably necessary. My first paragraph describes the DB-Kasparov problem domain, not the second. I think they were right not to try to solve the problems presented in my second paragraph, I would have done the same thing. If IBM's goal was to get the highest rated entity on ICC, they would have gone for the second paragraph. But it wasn't. You can't expect too much from a chess program. Nobody is saying that it does everything. A chess program is designed to solve problems in the domain of chess, the exact definition of this domain being affected by practical considerations which are not at all central. bruce
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.