>
> What I mean is that if human player H beats computer C1 65% of the
> time, and computer C2 also beats computer C1 65% of the time, then I
> would expect that H would be stronger than C2, especially if both C1
> and C2 are MC programs. If it is the case, then it would make it
> difficult to compare human scale to computer scale. But that is just
> my intuition.
>
> For instance, against computers, I estimate that Crazy Stone improved
> about 3 stones between this summer and now. But it clearly did not
> improve 3 stones on KGS. I vaguely remember that Sylvain also noticed
> that MoGo could beat GNU go with a 4-stone handicap, but was only 2
> stones stronger than GNU on KGS.
This happened in computer chess many years ago.    There was a period of
time when chess playing computers were relatively unknown but then
suddenly became very common.    This is anecdotal,   but it appears that
for a  certain amount of time the computers continued to improve, but at
the same time humans adapted very quickly to them.     Humans quickly
became educated.      They didn't become educated until computers
approached their playing strength.  

So there is a fairly significant ELO advantage to the human that has
experience playing computers.   I think it's very possible that this is
what you observed.    The advantage of a human however has limits.   You
are not going to beat a player 1000 ELO stronger just because you know
how he plays.    

Also, we don't see any serious discrepancy in computer vs human ratings
in chess although it was always imagined.   If you look in the sky you
can imagine interesting shapes,  but only because you are "looking" for
them.     Whenever we think we observe something based on a few data
points it's extremely subject to error.   Thus people often believed
that some given program might be 200 ELO stronger than other programs
but that it would translate into something very modest against
humans.    This NEVER turned out to be true - it was fantasy  based on
the continuing need to believe  that the improvements were just too good
to be true.  

A little common sense will tell you that this cannot be true but to a
very limited extent if any.    You would have to believe that the
in-transitive rubber band stretches to infinity.     Or you have to do
what Einstein did and regretted,  which is to impose an artificial
explanation such as some kind of constant that pushes this back at
higher strength levels.      The main point is that in computer chess
computers improved approximately 2000 ELO against humans over a couple
of decades or so.    But when it's claimed that ELO improvement against
other computers  is 4 times that,  you imply 8000 ELO points for
computer vs computer!     This is  clearly not the case.      That was
generally the claim I heard - 100 ELO improvement but it's someones
"belief" that against humans it's only about 25 ELO.      Nonsense.

I could see one things possibly happening however.   You might make a
real improvement that doesn't hit a human weakness that hard and at the
local limited  horizon it really may not translate to the same 
improvement against humans.    But intransitivity in playing strength is
like a rubber band.    You can only stretch it so far and it fights
back.     You are not going to get  A beats B 99% of the time,   B beats
C 99% of the time,  but  A cannot beat C.    Not unless you go way out
of your way to construct programs that have this behavior.       

I don't believe multi-dimensional playing strength  is much of an
issue.  It exists, but it's not severe.   You are not going to have to
worry that some 4 dan player will beat the world champion because he
happens to have a playing style that the world champion cannot handle.  
   Otherwise all rating/ranking systems would be useless.    And it
would be easy to  construct some really ludicrous cases of intransitivity. 


- Don












>
> Rémi
> _______________________________________________
> computer-go mailing list
> [email protected]
> http://www.computer-go.org/mailman/listinfo/computer-go/
>
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/

Reply via email to