HKA wrote:
vier wrote:
I don't understand. Difference in rating is defined by winning probability.
If a rating system is based on winning probability, then of course it is defined by such.
However, I think what John means is that the program is so precise, it can win against humans all the time. To such a rating program, that might lead eventually to a rating of 20 dan - but in another sense - could it give a 9 dan 11 stones?
To me it is like comparing Go Seigen and Lee ChangHo in their primes. It certainly looked as if Go Seigen was 11 dan - he could probably hold his own against 9 dans at 2 stones - certainly he proved he could at one stone. Lee Changho, on the other hand seemed precisely one or two points better than everyone else - not ranks but points.
John is questioning whether the program is an uber Go Seigen, or simply an infallible Lee Changho
All this discussion about KGS ranks is flawed. There is no evidence that the rating system can be extrapolated accurately so far from the mass of the existing data. Pros used to play with handicaps in the past and one handicap stone corresponded to approximately three dan levels. Thus, if Alphago Zero were actually 20 dan on the pro scale, human 9-dans would only take three or four stones
It might be interesting to see a trial in practice having Alphago Zero play Ke Jie with Ke taking handicaps to see just how big a handicap would be needed to give Ke a good chance of winning. Seems to me that it would not need such a big handicap. Maybe Alphago Zero wouldn't play, just resigning once the handicap stones were played.
Login
Register
FAQ
Search
Bill Spight
