Author: Kirill Kryukov
Date: 09:22:33 12/07/05
Go up one level in this thread
Hy Anthony, On December 07, 2005 at 11:29:59, Anthony wrote: >Yes, this doesn't surprise me too much. When there is a new entry, people that >get great results post immediately, and people that get bad results think they >did something wrong and redo their tests :) The same thing happened with >Ruffian. I also think that a big part of Rybka's strength is tactical, and >people posting blitz results will post earlier. I remain very impressed with >it, but it does appear that its been overhyped a bit. Of course, I'm OK with it >being overhyped if it increases interest in computer chess! Yeah, it was a bit overhyped, but 17 points more than nearest single-CPU engine (Fritz 9) is not bad at all either. And it really shows different chess than what we see from other engines. Of course, keep looking for the day when 64-bit version makes it into our main list. :-) >I am also curious which books you used for Shredder-Rybka and Fritz9-Rybka. We use a number of generic books, like Sedat's "perfect" books, Chesslib Elect, Harry Schnapp's books, etc. >anthony > >P.S. I really like your statistics page! Thanks! It is a team work and result of many discussions. >Some things I'd like to see: in the Pairwise page, how this performed relative >to expectations. Example: Engine X has rating 100 and Engine Y rating 150. >However, X scored 75% vs Y (+250 elo), so the value would then be +300 (X >performed 300 points better than expected vs Y). I am planning to do this actually, as soon as I figure the proper way to compute it. :-) >Also, on the "eval differences" page the raw eval difference isn't very >interesting IMO. For example, when Zappa is winning and Crafty is losing its >eval is usually +5 pawns higher or something. What I think would be more >interesting would be to sort the eval differences and then take the values at >25%, 50%, and 75%. So if the eval differences were 1 1 1 20 20 20 50 50 50 500 >500 500 we would get (1, 20, 50) rather than the arithmetic average which is >probably 100+. Actually there is missing explanations on my pages. Eval difference is taken only for pairs of consecutive moves, where eval on each of those move is not more than +- 900 cp. So the eval difference is not distructed by seeing mate, off-scale values from using TB, etc. Your suggestion is interesting too, but the simple truth is that as it is now, it is easy to understand for any reader. Anyone can open the page and realize what the table means. If I do your computation (similar thing was suggested by Uri before), most of people will simply not know how to connect those numbers to reality. Since I ignore the part of game where eval is more than 900 cp, and also I disregard all book moves and forced moves, the eval difference as it is now shows quite well how the engines actually differ in their evaluation, during the middlegame. Best, Kirill
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.