Tuesday
Mar202012
Ken Regan's Cheating Detection Program Profiled in the New York Times
Tuesday, March 20, 2012 at 9:59AM
The title is a bit too strong, as Ken Regan's program can certainly provide strong statistical support (or disconfirmation) for accusations of cheating, but not proof. Nevertheless, it's a useful tool with some nice fringe benefits (intrinsic performance ratings, for example, offer a comparatively objective metric for comparing past and present players). Read all about it in this New York Times profile.
tagged Ken Regan, New York Times, computers
Reader Comments (2)
I'd love to see a more in depth article about Ken Regan's results. The historical comparisons sound fascinating. I'd be curious if there was a difference in strength between chess generations if you start later in the game. Karpov may be the equal of the 40th best player today because his openings were worse but his late middlegame and endgame were better?
[DM: I'm sure Ken will give you a better and fuller answer, but as a start I'll offer a couple of quick points. First, the article talks about Karpov in the 1970s - he got better in the 1980s once Kasparov started pushing him. Second, while Karpov's rating eventually climbed into the mid- and upper-2700s, in the 1970s he was generally in the low 2700s (and at least at one point dropped into the high 2600s), which is where you'll find the #40 player in the world today.]
Thanks. Briefly, I've started a FAQ with supplementary information. Point 6. includes a link to three papers describing the work in detail, including my ongoing compendium of "Intrinsic Performance Ratings" based on analysis of the player's moves with my full (Multi-PV) model.
It computes Karpov's IPR in the 1978 win over Korchnoi as 2722, and he had 2721 while co-winning Montreal in 1979 (I have not entered individual data from tournaments yet). Merano 1981 = 2852, Moscow 1984-85 = 2762, Moscow 1985 = 2701, then he hit 2807 in London-Leningrad 1986 and 2838 in Seville 1987. (The paper also gives error bars; the last digits are not significant; the numbers will change as I upgrade the fitting procedure with more data, but a recent change made less than a +- 5 Elo difference; the overall Elo average for all WC matches since 1972 is 2743 while the "IPR average" is 2747, which is even closer than one would expect from two random takes of the same thing.)
If a volunteer will help run more Karpov performances from the 1970's in the depth-13 50-PV mode my full model requires (6-8 hours per core per game), I'll be grateful!