I doubt that much can be learned from comparing the overall length of rating
scales.
1: The large draw margin in chess compresses the high end of the chess Elo
range compared to go.
It takes a fairly large difference in skill for one very strong chess player to
win 65% against another very strong chess player, because they will draw a
large portion of their games.
In go, draws are absent, so it takes a smaller difference in skill for one very
strong player to score 65% against another very strong player.
If go ratings were pure Elo ratings, the high end would be more stretched than
the high end of chess ratings, just because of the absense of draws in go.
2: And then there is the fact that EGF is not pure Elo:
What if chess ratings were not based on winning percentages, but on 100 points
for every handicap step of "move odds"? (That is precisely what go ranks and
ratings are supposed to reflect)
Such a scale would probably "stretch" the range at the top and "compress" it at
the bottom ("move odds" being much more valuable to strong players than to weak
players).
I guess it would be possible to convert such handicap-based ratings to real Elo
ratings by collecting statistics, but the fit would probably be more complex
than a simple shift and constant scale factor.
But what else would be the learned from comparing the length of this scale to
the length of the normal Elo scale? Not much, I think. It's just a different
system.
So I agree with Christoph Birk that percentiles (when available), give better
absolute skill comparisons when comparing different games with different rating
systems.
Better at least than a rating conversion based on a simple shift and constant
scale factor.
Dave
________________________________
Van: [EMAIL PROTECTED] namens Don Dailey
Verzonden: wo 19-11-2008 20:59
Aan: computer-go
Onderwerp: RE: [computer-go] Re: Opportunity to promote ...
On Wed, 2008-11-19 at 10:24 -0800, Christoph Birk wrote:
> That should not matter much. The typical chess player should be
> "as strong" as the typical Go player and I also expect the strength
> distribution to follow similar lines.
Larry Kaufman, a chess Grandmaster and also an expert in many games
once told me that there are many more levels in Go than in chess.
For instance in chess, your rating can range from about 500 to 2800 or
so (roughly.) You can have a rating less than 500 of course and some
do, but 500 represents something like someone who just learned the rules
give or take. But for arguments sake let's say that 99% of all players
are well within a span of 3000 ELO points to be generous.
For 19x19 go, assuming we were to use ELO ratings the range would be
much higher according to Larry. So it could be more like 4000 or
5000, I don't know.
So I don't know for sure what you mean when you say the strength
distribution follows similar lines.
For comparison purposes maybe we need to identify the "median player"
somehow and extrapolate from there. But if you take the median for
both games, and assign them some fixed elo, I think you would find the
top GO players had ELO ratings hundreds of points higher than the top
Chess players.
There is yet no ceiling on how strong chess players can be either, so I
assume the same for GO, even more so. The top playing chess program
seems to be at least 200 ELO stronger than the best human and they
continue to improve every year. And it's still very clear that they
could improve a lot.
- Don
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
