Author: Michael Neish
Date: 19:41:57 01/18/00
Go up one level in this thread
>Walter expressed surprise at the result as well, since it differed from his >previous experiment. It just goes to show that a single experiment is hardly >conclusive. OTOH, considering the blitz ratings crafty gets on internet >servers, it is no slouch in that department. I keep intending to analyse the statistical variation that one would expect in the result from a series of games, and never get down to it. I wonder what Elo's book has to say about it (I don't have a copy). I sat down the other day and messed about with a few numbers, and found that if two computers of exactly equal strength play each other over 20 games, then, assuming there are no draws, you would expect a 50-50 score to occur only about 12% of the time, i.e., one in eight matches. That is, seven matches out of eight will be won by one or the other computer! And you could even expect a 12-8 score about 6% of the time, which is about one in seventeen. So an equal score will occur only about twice as often as a lopsided 12-8 score! I suspect this is old hat, but a lot of people seem to overlook this fact when apparently "strange" results appear. Does someone have a more thorough analysis of this kind of statistics? If not, then I'll have a go at it later this week, including draws and different expectations between White and Black, and send another post. I'm also interested in determining how many games are needed to establish, to 95% or 99% certainty say, which program is better. My argument is that statistics are probably good enough to explain the 15-9 score. What do others think? Cheers, Mike.
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.