Author: Stuart Cracraft
Date: 20:46:25 08/07/04
So after a lot of input from various people (thanks Alessandro, Gunther and the others), SEE appears to be up here and some test results in. The problem is that the theory looks great but the practice looks not so great. The version implemented uses SEE only in quiescence and only to avoid searching positions with negative see value. No SEE is used for move ordering, window-changing, or anything else anywhere. I tried those and they all gave much worse results for the second part of the test below. Two tests were done. Fixed depth and Fixed time. First, fixed depth of 8 ply in opening position: 2.02h without SEE: 1/ 2 e2e4 0.02 35 41 e2e4 2/ 4 e2e4 0.05 0 149 e2e4 e7e5 3/ 9 e2e4 0.07 17 450 e2e4 e7e5 f1c4 4/14 e2e4 0.08 0 2283 e2e4 e7e5 f1c4 f8c5 5/17 e2e4 0.12 17 7260 e2e4 e7e5 g1f3 b8c6 f1c4 6/23 e2e4 0.48 0 46344 e2e4 e7e5 g1f3 b8c6 f1c4 f8c5 7/26 e2e4 1.20 17 119943 e2e4 e7e5 g1f3 b8c6 f1c4 g8f6 b1c3 8/31 e2e4 6.15 0 699160 e2e4 e7e5 g1f3 b8c6 f1c4 g8f6 b1c3 f8c5 e2e4 6.15 0 699160 e2e4 e7e5 g1f3 b8c6 f1c4 g8f6 b1c3 f8c5 nps=113660 ha=17.00% q=57.0% bc=71% br=NaN% mp=108<>99 pawnx=0 recapx=0 qcheckx=27853 checkx=0 futilx=0 qfutilx=0 2.02i with SEE: Alpha=-250 Beta=250 Maxdepth=8 MaxTime=9999999 1/ 2 e2e4 0.00 35 41 e2e4 2/ 4 e2e4 0.01 0 149 e2e4 e7e5 3/ 6 e2e4 0.01 17 442 e2e4 e7e5 f1c4 4/10 e2e4 0.03 0 2101 e2e4 e7e5 f1c4 f8c5 5/12 e2e4 0.09 17 6248 e2e4 e7e5 g1f3 b8c6 f1c4 6/20 e2e4 0.37 0 37009 e2e4 e7e5 g1f3 b8c6 f1c4 f8c5 7/24 e2e4 0.91 17 94959 e2e4 e7e5 g1f3 b8c6 f1c4 g8f6 b1c3 8/29 e2e4 4.59 0 537608 e2e4 e7e5 g1f3 b8c6 f1c4 g8f6 b1c3 f8c5 e2e4 4.59 0 537608 e2e4 e7e5 g1f3 b8c6 f1c4 g8f6 b1c3 f8c5 nps=117185 ha=18.00% q=51.0% bc=63% br=NaN% mp=97<>99 pawnx=0 recapx=0 qcheckx=7616 checkx=0 futilx=0 qfutilx=0 q is the number quiescence-nodes/total-nodes. Iteration number remains the same due to fixed depth. Note how maximum depth reached (after the slash) decreases. Note how q decreased from 57 to 51. quiescent check extensions decrease by a factor of 4 or so. The final pv is the same. Total nodes searched drops 23%. Total time taken drops 25%. So all of this looks really good. Real pleasant. But now the fixed time result on 300 position problem set... Columns are version, % correct, correct/total totalseconds totalpositions averagepositionsperproblem/average seconds per problem/average nps per problem and the final non-zero number is the number of free check extensions in quiescence back into the main search. So thinking that this is great, I give it 300 time-limited problems (1 second per problem). It should search more deeply as a result and it should score well. But this was not to be. SEE is 2.02i. Pre-See is 2.02h. // 2.02i - 67% 201/300 267.55 57309292 191031/1/214199 0/0/1935723/0/0/0 Pre-See is 2.02h // 2.02h - 68% 205/300 267.99 62923440 209745/1/233764 0/0/3901373/0/0/0 So I am not too impressed yet. Is SEE, like so many things for real play, not as good at problems? I really don't want something that reduces tactical effectiveness!!! Anywhere! Stuart
This page took 0.01 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.