>From jrb9m@cms.mail.virginia.edu Tue Oct 10 19:07:30 2006 >From: "Jason R. Buczyna" >Subject: Rybka Game 2 done >To: Kenneth Regan Dear Ken, I have finished the Rybka 2.0c 32-bit analysis of game 2. I can definitely understand the reason for Danailov's suspicion: of the final 32 moves for Kramnik, 30 are matches, including 16 straight moves at one point. There are a number of forced moves, as well as ties, but even tossing those out gives a very high percentage, much higher than I'm used to. One would really have to analyse Kramnik's typical endgame play to see if this warrents action--though I don't think it's something I can do this week (though I'll start analysis on a game now, and do more in bits when I have time at the computer--I'll let you know which game I choose). Topalov's number for the last 32 moves is 78%, which is high, but more reasonable when you consider the forced moves and the ties. Out of all 48 moves, Topalov has 66.7% matching, just like with Fritz 9, while Kramnik has 79.2% matching. The text of the analysis is as follows; feel free to share it with whoever you deem appropriate: [Event "WCh"] [Site "Elista RUS"] [Date "2006.09.24"] ...etc. as in the file "KT2logJB.txt".] ------------------------------------------------------------------------ >From regan@cse.Buffalo.EDU Wed Oct 11 00:37:08 2006 >Date: Wed, 11 Oct 2006 00:37:08 -0400 (EDT) >From: Kenneth Regan >To: jrb9m@cms.mail.virginia.edu, regan@cse.Buffalo.EDU >Subject: Re: Rybka Game 2 done >Cc: regan@cse.Buffalo.EDU Dear Jason, Very Striking!---but explainable, I think. First some Qs: [Q's about his Rybka setup and timing info snipped.] ... ... Anyway, here's where my point about "deltas" and *prior probabilites* comes in. The short answer is: all but 9 of those moves were basically forced, plus 1 clear Kramnik blunder (Susan Polgar nailed it quick) and one move we didn't realize was 1/2-pawn worse than the better one. You don't have to be an endgame expert yourself you cam trust that Kramnik is an endgame expert...and look at Rybka's own diffs. The only ones under 0.73 (which == Fritz's criterion for "+/-") are: 33...Re7 +0.29 37...Qd1+ TIE (and most commentators felt checking first was wrong) 39...Qe4 +0.05 40...Qf5 TIE with 3 other moves, all 0.00 42...a5 +0.40 (obvious move to me) 44...a3 TIE (this time ...Rc7 is listed first by Rybka) 45...Bf8 +0.22 48...Rc1?! -0.52 NON-MATCH 49...Rf1? -2.63! NON-MATCH 53...Re1? TIE (correct ...Re3! listed 1st until 24 ply) 63...Re3 TIE, but conceptually obvious as well as best mate. Now to me the key to translating from engine evals to "prior probabilities" is to ask: At what difference in eval e(m1) - e(m2) would strong GMs in random tests prefer move m1 to move m2 68% of the time? I.e. what difference in eval (at *high* ply depth) represents 1 std. dev. in prior probability? I'm pretty sure it's a "delta" between 0.20 and 0.30. Thus I think 33...Re7 was a 70-percenter, 42...a5 an 85-percenter (based on evals; the move is clear to me as a player!). This leaves just 2 matches clearly within 1 st. dev., balanced against 2 definite non-matches, and then the TIEs. If you think the guy might choose the move Rybka lists first, even that splits down the middle! So while it's certainly striking when expressed by Hamming Metric, it "goes away" when a fair metric is used. This is a FANTASTIC example---of what (if no cheating) may have supplemented the bathroom behavior to create the illusion. I will post your stuff tomorrow---adding to my prelim files at www.cse.buffalo.edu/~regan/chess/fidelity/ (My Game 3 study cannot see how to reproduce the Bulgarians' %tages at all, unlike Game 6 whee I felt I'd explained their figures. It helps me that Fritz is *less* configurable than other engines...) Sincerely and really thanks, ---Ken R.