Author: Walter Faxon
Date: 01:04:39 01/07/03
Go up one level in this thread
On January 06, 2003 at 12:49:56, Graham Laight wrote: >Human chess is all about pattern recognition. Computers achieved their strength >though sheer speed. > >I doubt if there's time to do computer chess the correct way now - by the time a >pattern recogniser would be able to get up to the required strength, all games >will be a draw (this will happen at about 3500 elo - see >http://math.bu.edu/people/mg/ratings/Draws.jpg . Moreover, Rudolf Huber has >proven that there's no forced win in the first 30 moves. IMO, it is proven >beyond reasonable doubt (though not actually 100% proven) that there's no forced >win in ANY number of moves - chess is a draw). > >Perhaps those of us who believe that the best AI methodology is a database of >patterns from which to retrieve the closest matches to the patterns in the >current position should move on to Go - where the choices of moves on each turn >are sufficiently large to be entirely safe from the number crunchers in the >foreseeable future. > >-g Hi, Graham. Your 3 topics, restated: (1) It's too late for advanced AI pattern recognition techniques to be used in computer chess because such programs will never catch up to those using fast search with simple pattern recognizers. Or at least they won't catch up until all games are drawn anyway, as per topic (2). (2) Increased numbers of draws between humans as ratings increase strongly suggests that the game of chess is a draw, and computer chess will demonstrate this in a few years. (3) So AI gamers should tackle Go, because its state space is too large to tackle with anything like brute-force search. My take on them: (1) If we want a venue to explore advanced AI in chess, let's change the rules for a new series of competitions: "limited search" computer chess (LSCC). Humans search about 2 positions per second (Anand claims he searches 5). Well, our computers today are pretty dumb so let's start with a maximum 100 positions/second; this can be lowered later. The computer horsepower now used for searching would be applied to complex pattern recognition and other AI techniques. All program sources would be made public after each tournament, like the "RoboCup" robot soccer competitions today. This would serve two purposes: to prove nobody is searching faster than allowed, and to spread the wealth of knowledge. This would revitalize academic computer chess: everybody learns from everyone else and there is no point in competing unless you have a new idea or a much better implementation. Many more papers would be written. And any ideas developed could also be adopted by the fast searchers, as possible. (2) The best over-the-board chessplayers make a _lot_ of mistakes, particularly allowing small advantages to dissipate. The much lower rate of draws in top-level correspondence chess suggests that very minor advantages _can_ be exploited, given enough time, energy and imagination. A super-grandmaster computer might well prove that the advantage of the first move is enough to win. Why anyone should otherwise care whether chess is ultimately a win or a draw is another matter. (3) The extreme regularity of Go leaves open the possibility that a much simpler method of tackling that game might prove successful. But most people involved in this forum prefer to work on computer chess because we prefer chess. So: chess good; but AI better! We need brave visionaries to demonstrate this! -- Walter
This page took 0.01 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.