[computer-go] Win rate and handicap (was: Modern brute force search)

[David G Doshay <ddoshay@xxxxxxx>]

On 8, Nov 2004, at 4:13 PM, Christoph Birk wrote:

> drd@xxxxxxxxxxxxxxxxx wrote:
> > I don't know how to compare  this gap with GO.  If an omniscient chess
> > player could win 95% of it's  games against a Bobby Fischer, how would
> > that compare to "number of stones" in Go?
> About 3 stones,
I won't comment about omniscient and any particular player, but in
games between SlugGo and Many Faces, we play statistically even
at 4 stones. We played games between h = 1 and h = 6, and the
winning percentages move from Sluggo winning about 80% at h = 1
to SlugGo (extrapolated from data up to h=6) looking to loose about
80% at h = 7. These bounds are weak on the high side because we
did not play that many games at 5 or 6 stones.

So, between these two programs, the 95% win rate is 4 or 5 stones.

Here are the graphs put together by Doug Ridgway and his analysis:

***** snip *****
	http://dridgway.com/Go/sluggo_vs_MFG2.pdf
The 95% CI on the Result = 0
intercept would indicate a fair handicap lying in the range (3.0, 6.3).
Two possible concerns: there's maybe some lack of fit, and in any case
this uses the margin of victory, which is always kind of questionable. I
also did a fit to the outcomes alone: see
	http://dridgway.com/Go/sluggo_vs_MFG3.pdf
The circles and error bars are
the win rates and CI on those rates for each handicap (white wins the
tie), and the lines are the result of the fit. The range of p=0.5
intercept is (3.2, 6.1) -- not so different from the other fit.
***** end snip *****



Cheers,
David


_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/