Author: Dann Corbit
Date: 16:52:11 04/24/00
Go up one level in this thread
On April 24, 2000 at 18:58:41, Tom Kerrigan wrote: >On April 24, 2000 at 16:12:46, Robert Hyatt wrote: > >>On April 24, 2000 at 15:39:41, Tom Kerrigan wrote: >> >>>On April 24, 2000 at 14:15:16, Robert Hyatt wrote: >>> >>>>You can prove that blitz ratings are _not_ good predictors for standard ratings >>>>in every case. And _that_ was the issue. Bertil mentioned Genius and Nimzo >>>>as two cases. Ed added Rebel as a third. I had already added crafty as the >>>>first name mentioned... So I don't care if, on some occasions, a blitz rating >>>>does predict a standard rating. I care more about on some occasions, a blitz >>>>rating does _not_ predict a standard rating. >>> >>>So you got a list of people who don't care for blitz matches. Maybe they have >>>some anecdotal evidence showing that blitz matches are useless. But has anything >>>seriously been done to prove this? >>> >>>Experiments have shown that computer-computer match results can swing wildly, >>>even if you play 100 games. So if you really want to prove that blitz match >>>results are useless, it seems to me that you would have to play two matches of >>>at least 100 games each between two programs. And that would just prove that the >>>blitz results are useless for _those two programs playing against each other_. >>>It would take a tremendous amount of effort to prove that blitz results are >>>useless in general, even if you only take the top dozen or so programs into >>>consideration. >>> >>>>enough said... >>> >>>You're not allowed to end discussions here. >>> >>>-Tom >> >> >>Who said "blitz matches are _useless_???" I said "blitz matches do not show >>how the same two programs will do at longer time controls." Care to join the >>right conversation? And I can certainly end a discussion here from my end... > >My apologies. I thought you could make the logical "leap of faith" and tack on >the necessary "... for predicting performance at longer time controls." Unless the algorithms chosen are identical, it is clearly absurd to predict strength at longer time controls by using shorter time controls. Consider two programs A and B. A is dominated by algoritm P and B is dominated by alrogithm Q. Time for P is O(n*exp(n)) Time for Q is O(n*log(n)*exp(n)) P is much slower near the origin for whatever reasons. At longer time control, P will knock the stuffings out of Q, hence A will knock the stuffings out of B. But at shorter time controls, micro-opimizations may dominate.
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.