Author: Uri Blass
Date: 23:56:07 10/01/05
The idea is to reduce the history threshold of fruit. I looked at one of the games of fruit2.2 in WBEC and I found that fruit blundered because of history based pruning against Sjeng(you can download the game at the WBEC page in games 1-252 of the premier division) http://wbec-ridderkerk.nl/ The relevant position when fruit started to blunder was at move 44: I tested 3 options for fruit2.2 with clean 512 mbytes hash(I give only part of the analysis otherwise the post may be too long) first fruit default personality needs a very long time to find Rxd1 at depth 18 251: DeepSjeng X2-x64 - Fruit WCCC'05, WBEC11 Premier Division 2005 [D]7q/5pk1/2QP1np1/p7/4p3/4N1P1/P3KP2/3R3r b - - 0 1 Analysis by Fruit 2.2: 44...Rh2 45.Qb5 Qh3 46.Qe5 Qc8 47.Qxa5 Qe6 48.Rd2 Qh3 49.Qe5 Qc8 50.a4 Qc1 = (0.02) Depth: 13/38 00:00:13 10214kN 44...Qh5+ 45.g4 Qh3 46.Rxh1 Qf3+ 47.Ke1 Qxh1+ 48.Ke2 Qf3+ 49.Ke1 Qh1+ = (0.00) Depth: 13/38 00:00:14 11604kN 44...Qh5+ 45.g4 Qh3 46.Rxh1 Qf3+ 47.Ke1 Qxh1+ 48.Ke2 Qf3+ 49.Ke1 Qh1+ = (0.00) Depth: 14/38 00:00:17 14112kN 44...Qh5+ 45.g4 Qh3 46.Rxh1 Qf3+ 47.Ke1 Qxh1+ 48.Ke2 Qf3+ 49.Ke1 Qh1+ = (0.00) Depth: 15/44 00:00:26 22060kN 44...Qh5+ 45.g4 Qh3 46.Rxh1 Qf3+ 47.Ke1 Qxh1+ 48.Ke2 Qf3+ 49.Ke1 Qh1+ = (0.00) Depth: 16/44 00:00:40 34667kN 44...Qh5+ 45.g4 Qh3 46.Rxh1 Qf3+ 47.Ke1 Qxh1+ 48.Ke2 Qf3+ 49.Ke1 Qh1+ = (0.00) Depth: 17/46 00:01:09 61015kN 44...Qh5+ 45.g4 Nxg4 46.Rxh1 Nh2+ 47.Ng4 Qxg4+ 48.Kd2 Qf4+ 49.Kd1 Qf3+ 50.Kc2 Qxf2+ 51.Kb1 Nf3 52.Qc3+ Qd4 53.Qxd4+ Nxd4 54.d7 Ne6 55.Rd1 Nd8 56.Kc2 f5 57.Rd5 a4 58.Rb5 ² (0.43) Depth: 18/55 00:04:22 233637kN 44...Rh2 45.Qc3 Qh5+ 46.Kd2 Qb5 47.Kc1 Rh5 48.a4 Qd7 49.Ng4 Qxg4 50.d7 Rh8 51.d8Q Rxd8 52.Rxd8 Qf5 53.Qd4 Qg5+ 54.Kb2 Qf5 55.Qe3 Qe5+ 56.Ka2 ² (0.42) Depth: 18/58 00:15:55 798943kN 44...Rxd1 45.Kxd1 Qh5+ 46.Ke1 Qh1+ 47.Ke2 Qb1 48.a4 Qd3+ 49.Ke1 Qb1+ 50.Ke2 = (0.00) Depth: 18/58 00:17:35 893189kN (, 02.10.2005) second fruit with history threshold=50 can do it more than twice faster 251: DeepSjeng X2-x64 - Fruit WCCC'05, WBEC11 Premier Division 2005 7q/5pk1/2QP1np1/p7/4p3/4N1P1/P3KP2/3R3r b - - 0 1 Analysis by Fruit 2.2: 44...Qh5+ 45.g4 Qh3 46.Rxh1 Qf3+ 47.Kf1 Qxh1+ 48.Ke2 Qf3+ 49.Ke1 Qh1+ 50.Ke2 = (0.00) Depth: 13/44 00:00:19 14704kN 44...Qh5+ 45.g4 Qh3 46.Rxh1 Qf3+ 47.Kf1 Qxh1+ 48.Ke2 Qf3+ 49.Ke1 Qh1+ 50.Ke2 = (0.00) Depth: 14/44 00:00:22 17379kN 44...Qh5+ 45.g4 Qh3 46.Rxh1 Qf3+ 47.Kf1 Qxh1+ 48.Ke2 Qf3+ 49.Ke1 Qh1+ 50.Ke2 = (0.00) Depth: 15/44 00:00:30 25494kN 44...Qh5+ 45.g4 Qh3 46.Rxh1 Qf3+ 47.Kf1 Qxh1+ 48.Ke2 Qf3+ 49.Ke1 Qh1+ 50.Ke2 = (0.00) Depth: 16/44 00:01:00 52693kN 44...Qh5+ 45.g4 Qh3 46.Rxh1 Qf3+ 47.Ke1 Qxh1+ 48.Nf1 Qf3 49.g5 Ng4 50.Qc2 Qf5 51.Ng3 Qxg5 52.Qc3+ Qf6 53.Nxe4 Qxc3+ 54.Nxc3 Ne5 55.Ne4 f6 ² (0.61) Depth: 17/57 00:02:02 106753kN 44...Rh2 45.Qc3 Qh5+ 46.Kd2 Qb5 47.Kc1 Rh5 48.a4 Qd7 49.Ng4 Qxg4 50.d7 Rh8 51.d8Q Rxd8 52.Rxd8 Qe2 53.Qc5 Kh7 54.Kb1 ² (0.34) Depth: 17/61 00:06:41 335340kN 44...Rxd1 45.Kxd1 Qh1+ 46.Ke2 Qb1 47.a4 Qd3+ 48.Ke1 Qb1+ 49.Ke2 = (0.00) Depth: 17/61 00:07:32 382380kN Third fruit with no history pruning needs even less time(inspite of it I suggest testers to start by testing history pruning=50 and to continue testing reducing history pruning only if it performs better) 251: DeepSjeng X2-x64 - Fruit WCCC'05, WBEC11 Premier Division 2005 7q/5pk1/2QP1np1/p7/4p3/4N1P1/P3KP2/3R3r b - - 0 1 Analysis by Fruit 2.2: 44...Rh2 45.Qb5 Qh3 46.Qe5 Qc8 47.Qxa5 Qe6 48.Rd4 Nd7 49.g4 Qf6 50.Rxe4 Qxd6 = (0.01) Depth: 13/47 00:00:43 33997kN 44...Qh5+ 45.g4 Qh3 46.Rxh1 Qf3+ 47.Ke1 Qxh1+ 48.Ke2 Qf3+ 49.Ke1 Qh1+ = (0.00) Depth: 13/47 00:00:47 37549kN 44...Qh5+ 45.g4 Qh3 46.Rxh1 Qf3+ 47.Ke1 Qxh1+ 48.Ke2 Qf3+ 49.Ke1 Qh1+ = (0.00) Depth: 14/47 00:00:54 43952kN 44...Qh5+ 45.g4 Qh3 46.Rxh1 Qf3+ 47.Ke1 Qxh1+ 48.Nf1 Qf3 49.g5 Ng4 50.Qc2 Qf5 51.Ne3 Qxg5 52.Qxe4 Ne5 53.Qd4 ² (0.32) Depth: 15/47 00:01:37 81576kN 44...Rh2 45.Qc3 Qh5+ 46.Kd2 Qb5 47.Kc1 Rh5 48.a4 Qb6 49.d7 Rc5 50.Nc4 Qd8 51.Rd6 Rf5 52.Qe3 Kf8 53.Qh6+ Kg8 ² (0.31) Depth: 15/49 00:04:09 204284kN 44...Rxd1 45.Kxd1 Qh1+ 46.Ke2 Qb1 47.a4 Qd3+ 48.Ke1 Kg8 49.Qc8+ Kg7 50.Qc6 = (0.00) Depth: 15/49 00:04:28 222394kN 44...Rxd1 45.Kxd1 Qh1+ 46.Ke2 Qb1 47.a4 Qd3+ 48.Ke1 Kg8 49.Qc8+ Kg7 50.Qc6 = (0.00) Depth: 16/49 00:05:31 281103kN 44...Rxd1 45.Kxd1 Qh1+ 46.Ke2 Qb1 47.a4 Qd3+ 48.Ke1 Kg8 49.Qc8+ Kg7 50.Qc6 = (0.00) Depth: 17/50 00:07:01 365176kN (, 02.10.2005) Note that you may need the game to reproduce the exact analysis Note also that fruit blundered again in this game after Qh5+ g4 and played Nxg4 and I found that fruit with clear hash does not do that blunder and I guess that having knowledge in hash prevents some history based pruning and help Qh3 to fail low faster so Ng4 becomes better at depth 16. Uri
This page took 0.12 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.