Author: Miguel A. Ballicora
Date: 08:10:15 02/08/02
Go up one level in this thread
On February 08, 2002 at 08:18:45, Sune Fischer wrote: >On February 07, 2002 at 23:48:30, Dann Corbit wrote: > >>On February 07, 2002 at 18:35:26, Heiner Marxen wrote: >>>On February 07, 2002 at 16:46:21, Otello Gnaramori wrote: >>>>On February 06, 2002 at 17:38:51, Côme wrote: >>>>>Hi ! >>>>>Go here for an interesting project >>>>>http://wind.prohosting.com/chessweb/HTML/project.html >>>> >>>>It's very interesting but since it's based on pattern recognition and >>>>evolutionary algorithms it will take a time -> infinite to have some interesting >>>>result , comparable to the strength of today's classic engines. >>>> >>>>w.b.r. >>>>Otello >>> >>>While you may be quite right, the same is true for most amateur programmers. >>>Yet they start off and roll their own program. Why? >>>Because there is something to learn for them. Right? >>> >>>Dito for evolutionary methods applied to chess. There is still a lot to learn. >>>Also, evolutionary algorithms tend to be very sensitive to the details of >>>their setup, and so the failure of one or two experiments need not tell a >>>lot about the chances of the next experiment. >>> >>>For me at least it looks like a much too unattended topic. >>>I'd love to see some results, whether they are positive or negative. >> >>I agree. My gut feeling: It will be a dismal failure. However, I hope very >>strongly that I am wrong. And if it is a dismal failure, it is still not really >>a failure. Because we will know what *doesn't* work. That's one less thing >>that we'll have to try. Further, nobody will ever make an advancement in a new >>direction unless someone bothers to try it. > >I also don't think neural nets will be any good at playing chess, not at a high >level anyway. >What neural nets are good at, is to read peoples handwriting or identifying >(criminal) faces in a large crowd that's being photographed realtime by a video >camera. This is because their strength is _ignoring_ the details and picking out >a few features they recognize. >Chess is not like that, a knight in the center is not _always_ the best move, >and to my knowledge there is no "best" pawnstructure. > >The current algorithms do actually have some pattern recognition, checking for >doubled and connected pawns for instance. So using patterns as a small part of >the eval() is a good thing, but exact calculations will be needed. I believe >GM's also do exact calculations on most of their moves. > >-S. IMHO, pattern recognition in chess is not useful to evaluate a position in absolute terms. Not even humans do that. Pattern recognition would be useful to determine which plan is best, what moves deserve to be "extended" in the analysis and what kind of evaluation terms should a use. Too complex for a computer today, when they even do not know how to follow or create a plan. That is why only simple forms of pattern recognition may be succesful now. Regards, Miguel
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.