Author: Robert Hyatt
Date: 20:40:09 01/06/00
Go up one level in this thread
On January 06, 2000 at 19:47:10, Graham Laight wrote: >On January 06, 2000 at 17:20:44, Robert Hyatt wrote: > >>On January 06, 2000 at 10:43:29, Graham Laight wrote: >> >>>On January 06, 2000 at 10:23:46, Robert Hyatt wrote: >>> >>>>It is more than anecdotal. There is no contrary evidence at all, so far, other >>> >>>I don't agree - I think that the SSDF list represents "evidence", because they >>>have long experience of every level of play the computers have reached since >>>1984 or 1985. >>> >> >> >>What does the SSDF rating list have to do with whether a computer is at a GM >>level or not? You could add or subtract 400 points from every rating on their >>list, and things would still be just as valid according to the Elo formula. >>The 'spread' between two programs on the SSDF list is correct. The absolute >>value of the ratings are over-inflated. Or do you believe that a computer is >>really playing at 2700 and is in the top 10 in the world? >> >>I don't... > >If the 'spread' is correct, then the absolute values must also be right - >because the SSDF list is known to correlate well with FIDE ratings UP TO A >CERTAIN LEVEL (though it is admitted to be 20-30 points too high). where is this known? It was adjusted to approximate the Swedish federation ratings of a few human vs computer games 7 years ago. But nothing to do with FIDE type ratings... And the 'spread' and the absolute value are two different things. The spread predicts the outcome of a match between the two players. The absolute value of a rating is totally meaningless and is set by the initial rating given to the first group of players in the pool. The ratings could be 2500 or 25000 as far as that goes, as the value is meaningless. Once the pool stabilizes, the absolute values are _still_ meaningless. Only the spread between two players is meaningful. > >I think that you are saying that, relative to the FIDE ratings, the spread is >too great at the high end. I am not sure there. I just think that the overall numbers are way too high. I don't think any program is in the 2700 range at 40/2hr time controls. Not even close to that. But it _is_ an opinion, since we have so little data to support or contradict my opinion. But Ed is producing a little here and there. > >If it is true that the SSDF ratings correlate well with the FIDE ratings up to, >say, 2400 points (which probably is true), then what I think you are telling me >is that, for those computers above 2400 on the SSDF list, the gap between them >is too big, and that therefore the higher you get on the SSDF list, the more >overinflated the scores are, relative to human players. > >>>>than 'opinion polls'. Let's watch the Rebel games. That will be a reasonable >>>>guage... >>> >>>Certainly. Even better if the SSDF take up Ed's offer to test Rebel Century. >>> >>>-g >> >> >>That doesn't help a bit for the SSDF rating numbers. Their rating pool of >>players has nothing whatsoever to do with FIDE, so the ratings can't be compared >>at all. If they wanted, they could take rebel-10's eventual TPR as a real FIDE >>rating, then enter Rebel into the SSDF testing cycle, and when it finishes, >>reduce everyone's rating by X so that rebel's SSDF rating matches its TPR rating > >Agreed. > >>for the GM challenge matches. I think that X will be 200 points or more, IMHO. > >In my opinion, which is equally humble (of course!), Tiger's FIDE rating is >probably about 2660 - I don't think that this is quite in the top 10. > >-g Check out the FIDE list. It is surely top-25, which (IMHO) is not possible.
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.