Author: Sune Fischer
Date: 00:34:25 06/02/04
Go up one level in this thread
On June 01, 2004 at 23:44:02, Robert Hyatt wrote: >On June 01, 2004 at 20:46:45, Sune Fischer wrote: > >>On June 01, 2004 at 20:18:23, Robert Hyatt wrote: >> >>>>>If it can "use" it, yes... >>>> >>>>"use" it, certainly. >>>>The next time someone searches for that piece of data it will be there. >>> >>>Not the same thing. I don't use data crafty "learns". Crafty creates the data, >>>saves it, and uses it later, all without my assistance. That's the point... it >>>is something created by the program, for the program. >> >>Huh? >>So if the scientist helps the rat find through the maze the first few times, the >>rat is not learning? > >That's your words, not _mine_. I _always_ claimed that the rat is learning. And >the program is learning as well. Crude to be sure. But crude learning _is_ >learning... If you want to call the writing down of a friends phone number in your pocket book learning.... >>If I manually code new lines into the book, the book has certainly "learned" >>something new. > >Now we resort to semantics. Put your program on ICC, leave for 12 hours, and >see how it does. Mine will do fine because it _will_ learn, _by itself_ which >book lines are good and bad.. If you have to help yours "learn" then it is >worthless since you can't be everywhere it plays. I don't have that limitation, >so long as someone doesn't turn it off. Automated is nice, but what has that got to do with learning? >>Ever heard of "learning with a teacher"? :) > >Ever heard of learning alone? The "teacher" is what I chose to eliminate when I >started the book learning development in Crafty. I did it the "teacher" way far >too long. It is much more fun now. Automated is nice, but what has that got to do with learning? >> >>>>You think ICC games, I think home basement games. >>> >>>Same thing... >>> >> >>Absolutely not, ICC is worthless for systematic testing. >>You never know what your opponent is using, if he changed from last time etc. > >Ditto for basement tests as well. There are ways to do testing, but just >"playing games" is not a good one. Depends what you're testing. Suppose you change your nullmove scheme, are you going to look at 100 ICC games against random opponents and conclude "looks like this is worth 5 Elo"? Of course not, everything drowns in noise at that place. >> >>There is no way you can measure progress under those conditions, unless we are >>talking really huge progress. > >Absolutely wrong. I've been improving Crafty using _only_ ICC games for years >now. If you only look at 1's and 0's, you might have a problem at times, but >with so many games per week even that can work. But you get more than win or >lose, if you take the time to look. Your approace is unscientific and extremely ad hoc. Maybe a few Elo can be picked up like that but progressing must be much slower than it could have been. >You _are_ reading what I write? If you turn it off, I _will_ lose repeatedly. >That was my point... You keep forgetting the senario, which is a tournament in someones basement, a controlled environment where everybody has lame tricks disabled. -S. >>:) >> >>-S.
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.