Author: Robert Hyatt
Date: 07:35:49 03/03/98
Go up one level in this thread
On March 03, 1998 at 04:38:17, Ernst A. Heinz wrote: >On March 03, 1998 at 04:23:21, Ulrich Tuerke wrote: > >>Your results don't mean anything. You can't quesion hundreds of games >>which have been played by the SSDF by playing a couple of games. When >>you play 100 games, your result still has a statistical uncertainty of >>about 100 ELO points (order of magnitude). In order to derive a safe >>conclusion, you have to play that many games. I know it's a pity but >>that's life. >> >>Regards, Uli > >But what about the independence of samples (games) when the SSDF results >are full of doubles? > >I.e., the 14 French opening games Rebel vs. Fritz which Howard mentioned >seem to contain at least 10 doubles. Thence, their statistical relevance >shrinks significantly -- namely to that of 4 other games without >doubles! > >=Ernst= Not necessarily. IE what would you expect if You entered "X" in a tournament and it played me (me, Bob Hyatt, not me, Crafty) and I won. And the next week we meet again, same colors, and I play the same opening and win again? I call that stupid programming. "doubles" are something that a good program avoids if it is losing those games, And I think that statistically, it is something that probably evens out except for those programs that learn. But even a learner will play doubles since the "learning" isn't going to be shared (with crafty, it can be shared, but it is doubtful that anyone doing SSDF testing would take the time to do so). For those that complain about doubles, I'd say to the programmers "fix 'em" so they won't play a losing opening twice. It is not hard. Public code is available to show exactly how to do it, and it works well. I'd also suspect that games played on the servers are not of the same quality as SSDF games, because I doubt anyone is willing to manually operate two programs on servers and play 40/2hr time controls. You will likely see something like 15 15 as a *long* standard time control, and that is nowhere near 40/2..
This page took 0.01 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.