Author: Vasik Rajlich
Date: 01:20:51 03/18/04
Go up one level in this thread
On March 17, 2004 at 21:22:37, James Swafford wrote: >On March 17, 2004 at 21:19:45, James Swafford wrote: > >>Just what is a good eval worth? >> >>I ran Crafty on an AMD XP 2000+ (about 1.5 ghz) on ICC with a stripped >>down eval, then with it's regular eval to find out. Check out the >>summaries below (the top one is for stripped eval, of course). >> >>By "stripped down", I mean that the eval counted only material, some >>very crude pawn heuristics (doubled pawn, passed pawn, isolated >>pawn), and used pawn and knight piece square tables. That's it. >> >>I'm a bit surprised that the difference isn't a bit more; it's >>less than 200 points for blitz and bullet. I strongly suspect >>the difference would've been much greater for standard games, >>but I have no data to support that. >> >>Anyone done similar tests? >> >> >> cnt | avg | avg | base_secs | inc_secs | >>description >>-----+-----------------------+-----------------------+-----------+----------+------------- >> 271 | 2313.6531365313653137 | 1731.1623616236162362 | 120 | 1 | Br >> 2 | 2431.0000000000000000 | 1031.0000000000000000 | 300 | 3 | bu >> 673 | 2409.4680534918276374 | 1654.3239227340267459 | 300 | 3 | br >>(3 rows) > > >I should've explained here that the first "avg" column is Crafty's >average rating, the second is Crafty's opponent's avg rating. >Obviously I did not restrict ratings: I just did a >seek 2 1 r; seek 5 3 r > >-- >James > > >> >> cnt | avg | avg | base_secs | inc_secs | >>description >>-----+-----------------------+-----------------------+-----------+----------+------------- >> 246 | 2476.1422764227642276 | 1644.9471544715447154 | 120 | 1 | Br >> 2 | 2436.0000000000000000 | 1845.0000000000000000 | 120 | 1 | Bu >> 567 | 2583.1869488536155203 | 1653.4673721340388007 | 300 | 3 | br >> >> >>-- >>James Interesting. It would also be interesting if the eval was slightly richer, and included: 1) Some piece-square tables for bishops, rooks, queens, and kings. Of course generally it's good to centralize all of these pieces, and if you have nothing the eval has absolutely no way to distinguish at all between piece placement. 2) Some slightly more sophisticated material values. A piece should be worth more than 3 pawns, two pieces > rook and pawn, and there should be a bishop pair bonus. Of course the experiment could go on forever, but I suspect that the above 2 may account for a very large chunk of the rating difference which you measured, and they are both ultra-cheap and ultra-easy to compute. It would also be interesting to see if the difference was less in computer vs computer play, as I suspect. Cheers, Vas
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.