Author: Vasik Rajlich
Date: 04:54:09 06/13/05
Go up one level in this thread
On June 11, 2005 at 21:05:13, Eelco de Groot wrote: >On June 11, 2005 at 15:42:31, Terry Giles wrote: > >>On June 11, 2005 at 09:48:06, Yar wrote: >> >>>[D] r1r1q1k1/6p1/p2b1p1p/1p1PpP2/PPp5/2P4P/R1B2QP1/R5K1 w - - >>>Despite Shredder 9 UCI was able to achieve ply 29, it wants to play Qb6: >>> 16/24 0:00 +1.75 1.Qb6 Rd8 (260.994) 397 >>> 17/41 0:24 +1.90 1.Qb6 Qf8 2.axb5 axb5 3.Qxb5 Rxa2 4.Rxa2 Qe7 5.Be4 Qc7 >>>6.Kf1 h5 (10.096.589) 404 >>> 18/42 1:36 +1.91 1.Qb6 Rd8 2.Be4 bxa4 (37.737.331) 393 >>> 19/50 4:00 +1.66-- 1.Qb6 Rd8 (90.372.017) 375 >>> 19/50 4:43 +1.50 1.Qb6 Qe7 2.axb5 Rab8 3.Qxa6 e4 4.Re1 Qd8 5.Raa1 Ra8 6.Qb7 >>>Rab8 7.Qa7 Bc7 8.Qa2 Be5 (105.014.911) 369 >>> 20/44 7:51 +1.37 1.Qb6 Qe7 2.axb5 Rab8 3.Qxa6 e4 4.Qa7 Qe5 5.Qe3 Re8 6.b6 >>>Qxf5 7.Ra7 Qxd5 8.Rd1 Qe5 9.Bxe4 Qh2+ 10.Kf1 Rbd8 11.Qd4 Qf4+ 12.Bf3 Qg3 >>>(169.399.949) 359 >>> 21/50 13:15 +1.46 1.Qb6 Qe7 2.axb5 Rab8 3.Qxa6 e4 4.Qa7 Qe5 5.Qe3 Re8 6.b6 >>>Qxf5 7.Ra7 Qe5 8.Ba4 Re7 9.Bc6 Qh2+ 10.Kf1 f5 11.Rxe7 (283.619.295) 356 >>> 22/50 20:25 +1.21-- 1.Qb6 Qe7 (423.113.178) 345 >>> 22/57 41:37 +1.04 1.Qb6 Qe7 2.axb5 Rab8 3.Qxa6 e4 4.Qa7 Qe5 5.Qe3 Re8 6.b6 >>>Qxf5 7.Ba4 Re7 8.Rd1 Qh5 9.g4 Qg6 (824.742.918) 330 >>> 22/57 50:53 +1.05++ 1.Kf1 (1.003.794.712) 328 >>> 22/57 59:22 +1.05 1.Kf1 Rd8 (1.163.032.721) 326 >>> 23/51 68:13 +0.80-- 1.Kf1 Rd8 (1.338.053.366) 326 >>> 23/56 110:11 +0.41 1.Kf1 Rcb8 2.axb5 (2.217.178.132) 335 >>> 23/61 119:37 +0.42++ 1.Qb6 (2.419.703.465) 337 >>> 23/61 126:54 +0.77++ 1.Qb6 (2.577.146.539) 338 >>> 23/61 152:21 +1.22 1.Qb6 Qe7 2.axb5 Rab8 3.Qxa6 e4 4.Qa7 Qe5 5.Qe3 Re8 6.b6 >>>Qxf5 7.Rd1 Qg6 8.Ba4 Re7 9.Bb5 f5 10.Bxc4 f4 11.Qd4 f3 12.Re1 (3.109.727.742) >>>340 >>> 24/61 201:40 +1.04 1.Qb6 Qe7 2.axb5 Rab8 3.Qxa6 e4 4.Qa7 Qe5 5.Qe3 Re8 6.b6 >>>Qxf5 7.Ba4 Re7 8.Bc6 Qe5 9.Ra8 Qh2+ 10.Kf1 Kh7 11.Rxb8 Bxb8 (4.079.766.362) 337 >>> 25/65 385:56 +0.79-- 1.Qb6 Qe7 (7.839.484.758) 338 >>> 25/66 537:43 +0.69 1.Qb6 Qe7 2.axb5 Rab8 3.Qxa6 e4 4.Qa7 Bc7 5.Ra6 Qe5 6.Qe3 >>>Re8 7.Bxe4 Qxe4 8.Qxe4 Rxe4 9.d6 Bd8 10.g4 Kf8 11.Ra8 Bb6+ 12.Kg2 Re2+ 13.Kg3 >>>Re3+ 14.Kf4 (11.114.015.920) 344 >>> 26/66 850:02 +0.59 1.Qb6 Qe7 2.axb5 Rab8 3.Qxa6 e4 4.Qa7 Bc7 5.Ra6 Qe5 6.Qe3 >>>Re8 7.Rd1 Qh2+ 8.Kf1 Bf4 (17.839.249.763) 349 >>> 27/69 1686:56 +0.52 1.Qb6 Qe7 2.Be4 Rcb8 3.Qe3 Qd8 4.Kh2 a5 (35.239.538.915) >>>348 >>> 28/68 2612:01 +0.51 1.Qb6 Qe7 2.axb5 Rab8 3.Qe3 axb5 4.Kf1 Qd8 5.Ra6 Ra8 >>>(55.003.447.943) 350 >>> 29/74 4964:35 +0.51 1.Qb6 Qe7 2.axb5 Rab8 3.Qxa6 e4 4.Qa7 Bc7 5.Ra6 Qe5 6.Qe3 >>>Re8 7.Bxe4 Qxe4 8.Qxe4 Rxe4 9.d6 Bd8 10.Rc6 Kf8 11.Kh2 Re5 12.Rxc4 Rexb5 >>>(104.701.273.354) 351 >>>best move: Qf2-b6 time: 5572:50.890 min n/s: 350.472 CPU 96.3% n/s(1CPU): >>>363.937 nodes: 117.187.868.493 Interesting. So 36. Qb6 fails low on ply 22 after 20 minutes (to +1.04), and then again on ply 25 after 385 minutes (to +0.69). Probably Deep Blue was able to reach that first point in the main (3 minute) search, at which point some small evaluation differences can explain the preference for 36. axb5. I've seen a number of positions like this, where a king attack is understood via a really huge search. Vas >> >> >>Who knows? >> >>Apparently some current chess programs/engines do choose to play Be4. Kasparov's >>concern revolved around as to how Deep Blue could prefer this move (after about >>15 mins thinking time) over the clear material winning move Qb6. Also after >>playing this and other good moves why did Deep Blue allow Kasparov a simple draw >>by perpetual check? (which however he sadly missed). It was this apparent >>inconsistancy that bugged Kasparov throughout the rest of the match and made him >>believe that there had been some human intervention. >> >>Terry > > >Finding 37.Be4 seems to be doable for Pro Deo 1.1 for instance, although not >very easy. > >[D]r1r1q1k1/6p1/3b1p1p/1p1PpP2/1Pp5/2P4P/R1B2QP1/R5K1 w - - 0 1 > >Output with Athlon 3200+ 200 MB HT default settings Pro Deo 1.1 > >00:00:00.2 0,10 1 7 Rxa8 Rxa8 >00:00:00.2 1,16 1 9 Ra7 >00:00:00.2 0,57 2 158 Ra7 Rxa7 Rxa7 >00:00:00.2 0,42 3 508 Ra7 Rxa7 Rxa7 >00:00:00.2 0,36 4 1497 Ra7 Rxa7 Qxa7 Rc7 Qb6 >00:00:00.2 0,39 4 2266 Ra6 Rxa6 Rxa6 Qd7 >00:00:00.2 0,49 4 9117 Qb6 Rxa2 Rxa2 Qd7 Ra7 >00:00:00.2 0,53 5 12044 Qb6 Rxa2 Rxa2 Rd8 Ra7 e4 >00:00:00.2 0,82 6 24958 Qb6 Rxa2 Rxa2 Bc7 Qe6+ Qxe6 >00:00:00.2 0,46 7 42875 Qb6 Rxa2 Rxa2 Bc7 Qb7 Kh8 Ra7 >00:00:00.3 0,68 8 100244 Qb6 Rxa2 Rxa2 Bc7 Qe6+ Kh8 Ra6 Ra8 Rxa8 Qxa8 >00:00:00.4 0,64 9 179427 Qb6 Rxa2 Rxa2 Bc7 Qe6+ Kh8 Be4 Rd8 Ra6 Ra8 Rxa8 Qxa8 >Qe7 Qa1+ Kh2 >00:00:00.7 0,64 10 498239 Qb6 Rxa2 Rxa2 Bc7 Qe6+ Kh8 Be4 Rd8 Ra6 Ra8 Rxa8 Qxa8 >Qe7 Qa1+ Kh2 >00:00:01.2 0,72 11 1146205 Qb6 Rxa2 Rxa2 Bc7 Qe6+ Kh8 Ra6 Rd8 Be4 Rc8 g3 Rd8 >Qxe8+ Rxe8 >00:00:01.8 0,72 12 1784681 Qb6 Rxa2 Rxa2 Bc7 Qe6+ Kh8 Ra6 Rd8 Be4 Rc8 g3 Rd8 >Qxe8+ Rxe8 >00:00:04.8 0,83 13 5186423 Qb6 Rxa2 Rxa2 Bc7 Qe6+ Kh8 Ra6 Qd8 Kh2 Qf8 g3 Qe8 Be4 >Ra8 Rxa8 Qxa8 >00:00:07.5 0,83 14 8355674 Qb6 Rxa2 Rxa2 Bc7 Qe6+ Kh8 Ra6 Qd8 Kh2 Qf8 g3 Qe8 Be4 >Ra8 Rxa8 Qxa8 Bg2 Qa1 Qc8+ Kh7 Qxc7 Qxc3 >00:00:20.7 0,81 15 23563148 Qb6 Rxa2 Rxa2 Ra8 Rxa8 Qxa8 Qxd6 Qa1+ Kh2 Qf1 Qb8+ >00:00:43.8 0,37 16 49556486 Qb6 Rxa2 Rxa2 Ra8 Ra6 >00:04:35.0 0,48 16 301499599 Be4 Rcb8 Ra6 Qc8 Qa2 >00:06:33.7 0,48 17 433939286 Be4 Rcb8 Ra6 Qc8 Qa2 >00:07:46.0 0,48 17 518483534 Qb6 >00:09:19.8 0,63 17 626589998 Qb6 Rxa2 Rxa2 >00:12:48.7 0,35 18 869864817 Qb6 Rxa2 Rxa2 Ra8 Ra5 Bxb4 cxb4 c3 Rxb5 >00:17:21.0 0,55 18 1174471267 Be4 Rcb8 Ra6 Qc8 Qa2 >00:40:46.9 0,54 19 2781409839 Be4 Rcb8 Ra6 Qc8 Qa2 Rxa6 Qxa6 Qxa6 >01:38:47.1 0,55 20 2458441397 Be4 Rcb8 Ra6 Qc8 >06:07:09.4 0,58 21 3628836035 Be4 Rcb8 Ra6 > >If the score of +0.58 from Pro Deo at 21 ply for 36.axb5 axb5 37.Be4 can be >directly compared to Shredder's +0.51 for a direct 36.Qb6, then Deep Blue could >arguably have computed 36.axb5 as about equal to 36.Qb6. > >However as far as I know it these depths, 23 plies or more, were not achievable >by the IBM team, even including Deep Blues hardware search. I personally think >that playing 36.axb5 may point to instability in the Deep Blue search. It is not >for nothing that Donninger says he lets the FGPA chips do as little as possible >compared to the software search. It is just very hard to tune hardware :). That >this proposed instability created some Anti-GM like moves like axb5 instead of >bigger blunders was IBMs luck but instability in the search may well have been >one of the reasons to put Deep Blue into the mothballs as quickly as possible >after the match. IBM would not have been so lucky twice. > >Just my two cents for what could have happened, I think Deep Blue was lucky to >win this match and it is a shame Kasparov did not get a rematch, from a sporting >point of view. > > Eelco
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.