Author: walter irvin
Date: 19:03:08 08/09/00
Go up one level in this thread
On August 09, 2000 at 14:55:17, Dann Corbit wrote: >On August 09, 2000 at 12:47:01, walter irvin wrote: > >>why is not more time spent on better learning for programs???? > >Because nobody has figured out an ideal way to do it yet. Your program finds a >good move at 40/2 and salts it away. Two years later, when your hardware is 4x >faster, that good move is now a poor move, because the computer may easily find >a better one. > >>while there is a >>vast number of legal positions ,the number of positions that a strong program >>will encounter is much smaller . > >I have a little over 100 million positions in my game database. I suspect if I >were able to download all the chess games from FICS and ICC it would double or >triple that. At 12 minutes per position it represents a significant investment >in time to analyze them all. Not saying that is a bad idea, but you'd have to >be off your rocker to try it. 12 mins a position you dont have to analyze anything .what you would have to do is figure out how many individual seperate winning positions you have .you only have to know if you won the game or lost the game .games won on time would not count .because you can win in a lost position that way . >;-) > >>plus just not playing a position when there is a >>negative score on that position doesn't really work because :sometimes the bad >>move is 5 or 6 moves back, but takes that many moves to be realized .sometimes >>the losing move is not easily know .the only way to go is to record winning >>positions and delete losing ones > >Define winning position. If I lose sometime sometime after playing 1. e4 has 1. >e4 now become a losing position? there should be a very small and short oppening book that that is never erased but it should only go 4 or 5 moves deep . > >>.that way all you are ever left with are the >>winning positions that can stand the test of time .why is this not done .instead >>a programmer will spend hours trying to fine tune a evaluation that may play >>better in some position and woarse in others , at least it plays different >>???????? > >Because while learning is a good idea, it is probably not a great idea. What >programs would have learned 10 years ago would be almost completely valueless >today. Take a 64 CPU alpha machine and install a SMP crafty on it, and the >learning done on a single CPU machine suddenly has become laughable. depends on the competition that crafty played .remember all the winning position will be stored even the ones by the GM . if a position is shown to be bad it is simply removed ,but at the same time a whole other set of winning positions are recorded .the bottom line is a bad position always gets replaced .a good position never gets replaced .i mean if crafty is clearly winning a game it will be very hard for even a GM to get the upper hand . also in a case like above .put a 386 crafty vs 64 cpu alpha make play 2000 games .then take a separate 386 crafty and make it play the one that learned vs 64 cpu .the one that learned would destroy the one that had not played the 64 cpu .learning is only going to be to the level of the competition that goes for human or computer . > >If there were easy and workable solutions to problems like this, programmers >would implement them. There aren't any.(YET) dont mean you should not seek them out . > >Learning (in general) is a pretty good idea and most advanced programs do some >learning. But it's not the panacea you think it is. NOT IN CURRENT FORM your right . > >Just a minor suggestion: >Judicious use of the shift key, space-bar, and punctuation marks might make your >posts a bit more readable. >IMO-YMMV
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.