Author: Ed Schröder
Date: 20:09:17 08/06/98
Go up one level in this thread
>Here are (just) 3 examples of the so called "holes" (as chess programmers >call it) in the selective search part of a chess program, in this case Rebel. >3rkb1r/p2nqppp/5n2/1B2p1B1/4P3/1Q6/PPP2PPP/2KR3R w kq - bm Rxd7 >r1r2b2/pb4pk/1pB2p1p/3P1P2/8/PP2q1P1/2Q1P2P/RN2R2K b - - bm Rxc6 >r2r2k1/Bpq2pbp/4p3/n3Pp2/5P1N/2P1Q3/P1P3PP/R4RK1 b - - bm Rxa7 >If you test these positions with Rebel you will notice it takes Rebel a long >time to solve these (simple) positions. This because of the too small (is >risky) selective search. I even think (not tested) that Rebel Decade does >a better job in these 3 particular positions :)) Fixed in Rebel10, it only >takes a few seconds now. >If this is worth the pretty huge investment of 30%, only time can tell but >I think it does.- Ed - >But - as You know, too - Rebel9 with sel=3 does a lot better job, e.g.: >3rkb1r/p2nqppp/5n2/1B2p1B1/4P3/1Q6/PPP2PPP/2KR3R w kq - bm Rxd7 ce 418; pv >00:25 > 8.00 4.18 Rd1xd7 Nf6xd7 Bg5xe7 Bf8xe7 Rh1-d1 a7-a6 Bb5xd7+ ; c0 Analysis by >Rebel 9.0; c1 30 seconds per move; c2 Pentium at 460 Mhz; >r1r2b2/pb4pk/1pB2p1p/3P1P2/8/PP2q1P1/2Q1P2P/RN2R2K b - - bm Rxc6 ce 396; pv >00:27 10.00 3.96 Rc8xc6 d5xc6 Ra8-c8 Nb1-d2 Rc8xc6 Qc2-e4 Rc6-c7 ; c0 >Analysis >by Rebel 9.0; c1 30 seconds per move; c2 Pentium at 460 Mhz; >r2r2k1/Bpq2pbp/4p3/n3Pp2/5P1N/2P1Q3/P1P3PP/R4RK1 b - - bm Rxa ce 240; pv 00:49 >9.00 2.41 Ra8xa7 Rf1-b1 Ra7-a8 Nh4-f3 Na5-c4 Qe3-f2 f7-f6 ; c0 Analysis by >Rebel 9.0; c1 30 seconds per move; c2 Pentium at 460 Mhz; Try "Brute Force = ON" and you will see even faster solution times. These so called "holes" in the selective part (or null-move) are the price a programmer has to pay for using selective search (null-move) and these holes exist in every (selective) chess program, only the (type of) positions are different. There is simply no way to replace "brute-force" with "selective search" without the risk you prune moves (complete variations) from the tree search which should not have been pruned. So a chess programmer takes a calculated risk using selective search or null-move. The alternative is to go back to "brute force" where simply nothing is pruned. However doing so a chess program will horrible slow down in ply-depth. Speaking for Rebel only I estimate a 7 ply search will slow down with a factor of 4-5, a 10 ply search will slow down with a factor of 8-10 and so on... Taking this into consideration (and after having tested this of course) it is simply wise to use "selective search" after all, as it certainly is good for 100-150 elo points (and maybe more) if you have a proper selective algorithm. Every year at least 1-2 months is spent to improve selective search, a never ending story because of its importance. >(could have done even better in pure DOS, but...) >Then it could be interessting to see the equivalent figures for the comming >Rebel10, if You got the time. And further - would You recommend (in general) >the sel=3 for Rebel9 on a fast machine? I can not recommend "sel=3" being stronger than the default setting. But "sel=3" isn't a bad choice and "sel=4" / "sel=5" are also good options, maybe 3/4/5 come close to Rebel's standard setting. Solution times (PII-266) of R10 for the 3 positions... Rxd7 0:08 Rxc6 0:02 Rxa7 0:08 Still not perfect but definitely better than R9. - Ed -
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.