Author: Andy Serpa
Date: 22:41:17 10/11/00
Go up one level in this thread
I've have toyed with this problem with some semi-interesting results. The main problem to overcome with the patten recognition approach is that GA's and NN's and CS's (Classifier Systems) are made to generalize, when of course chess positions are all about exceptions to the rule. If you start talking about "fitness landscapes", you will find chess positions have a HUGE degree of "epistasis" (I believe that is the term). That is, change one tiny little thing about the pattern, and you might have a totally opposite situation. So I believe the patterns being recognized have to be "mini-patterns" within the position, with some sort of overseeing algorithm to determine which ones are important or are out-weighed by others. I've never been interested in making a traditional brute-force type program simply because it seems boring. I'm more interested in the problem-solving that in solving the problem. I would like to see more approaches that attempt to unlock the secrets of chess. After all, when you search ahead "manually" 10 plys or so, aren't you just revealing the future potential of the current position? Are chess positions one-way ratchets like the best encryption algorithms, or can they be broken down and decoded through some other way? Ideally, a chess program shouldn't have to search ahead at all, but simply be "able to tell" what it happening, or at least if the position is won for one side or the other... Andy Serpa chessmad@chessopolis.com http://www.chessopolis.com
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.