Author: Dave Gomboc
Date: 16:10:00 02/13/03
Go up one level in this thread
On February 09, 2003 at 01:04:41, Antonio Dieguez wrote: >On February 08, 2003 at 06:28:00, Dave Gomboc wrote: > >>On February 08, 2003 at 03:30:34, Peter Kappler wrote: >> >>> >>>He has some good ideas. >>> >>>http://www.worldchessrating.com/521683950.html?529637011233717 >>> >>>-Peter >> >>The one thing that always bugs me about Sonas's stuff is that he takes the data, >>does some fitting on it, then says "see, this fits the data better!". Well, of >>course it does! > >yes and that's good. "Some fitting", you are being despective with the way he >does it. I don't see anything so weird or so senseless. > >Let imagine if I only tune the K factor with real data instead of no data. If I >do that, I wonder who can't believe that's no good. And so on with a few other >things. If there were a sound theory behind dictaminating wich should be the >best k then it is other thing though. > >> What he should do is have training data that is *distinct* from >>his testing data, e.g. tune his formulae on results from (for example) >>1990-1995, then test on (for example) 1996-2000. > >and after test and correct, end up with the same thing... just kidding > >> Tuning on 1990-2000, then >>testing on 1990-2000 is bogus. > >he is tuning on 1990-2000 and betting it will work better for other periods. Yeah, he's betting on it, but he can't show it. Why? Because he has no other data! If he was being scientific, he's split his training data from his test data, then he'd be able to show that his technique has predictive power into the future. Dave
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.