Hi,
This is a quote of a post between Douglas Zare and Frankl
Berger (author of BG Blitz) in RGB I thought might interest the group. Zare is
the one who suggested using unbiased Variance Reduction in matches:
> > Yes, I've seen it too. Why they havn't published it? Only speculation,
> > but
> > let's assume, that Snowie 4 has an equity of 0.03 ppg cubeless against
> > GNU (I believe it is less, but let's take that as an upper bound), how
> > many games you have to play, to make sure, that Snowie 4 is that much
> > better? Several thousands!! Doing thousands by hand is impossible and
> > having let's say 100 5pt matches proves nothing.
>
> A hundred matches means a lot if you use unbiased variance reduction.
Is variance reduction usable at a matches too? How do you measure the
luck in a match? (and the bots being at the same time the thing to be
measured and the tool to measure seems to be a problem to me..). And
AFAIK has no one till now developed the formula's to do this, so I
still feel hundred matches are not enough to *prove* anything. Having
a standard deviation of about 3/sqrt(number_of_games), you need pretty
high numbers for a 95% confidence interval...