Author: Johan de Koning
Date: 15:26:17 09/26/03
Go up one level in this thread
On September 25, 2003 at 04:55:52, Tim Foden wrote:
>On September 25, 2003 at 03:24:27, Johan de Koning wrote:
[...]
>>But I don't think symmetry gains much, since the root moves in the "other half"
>>spawn small subtrees. Maybe it helps a lot if you apply *all* translations and
>>mirrors to the transposition table.
>
>Yes, this is what I've done. I generate 2 hash keys (horizontally reflected),
>and load/store in the transposition table using the lowest valued key.
>
>In my tests it made nearly a 2x speedup.
Ok, I wasn't aware you exploited symmetry *in* the TT.
Now I'm tempted to try it too ...
And maybe stripping empty files, or even Uri's idea of stripping all files that
won't play a part in a position's future. That will surely f'up move ordering.
Hmmm, maybe it's time to let go.
>I've also tried with 4 hash keys, including black/white symmetry, but I got
>strange results, which were only marginally better in some cases, but mostly
>worse.
>
>I've yet to figure out if this was a bug in the implementation, or a flaw in the
>theory that this would be a correct thing to do.
That's probably because the mirrored bounds are mostly useless.
For example, consider White has found a root move with a +0.10 score, then:
1. d4 d5 search says wtm : w <= 0.10, stored, then:
1. d3 d5 2. d4 TT says btm : b <= 0.10, while b >= -0.10 needed to cut.
[...]
>>For the record, the outputs above come from an XP2200(1800MHz) with a 16M entry
>>TT (with experimental replacement scheme). Since the node counts peak in their
>>late 20s, I'm expecting 27 or 29 to be the heaviest iteration on the initial
>>8-8 position. It's running already and has produced a 27 ply PV in less then
>>2 hours. Full report will follow.
>
>I'll look out for it. I've stopped working on it for the moment, as real work
>is interfering. :)
Yeah, real work sucks.
Anyway here's the result after 1.27 day and 1.27e11 nodes.
[D]8/pppppppp/8/8/8/8/PPPPPPPP/8 w - -
6936.1 7851M D27 +5 : d4,d5 f4,f5 a4,a5 b3,c6 b4,axb4 a5,h6 h3,b3 cxb3 ...
23909. 27186M D29 +0 : d4,b5 e4,g5 c3,a5 f3,c6 g3,d6 e5,dxe5 dxe5,h5 h4 ...
33271. 37816M D29 +5 : c4,f5 f4,c5 d3,d6 e3,e6 b3,b6 g3,g6 a3,d5 g4,h5 ...
54588. 62497M D31 +9963 : c4,f5 d4,a5 f4,a4 b4,axb3 axb3,c6 b4,b6 d5,cxd5 ...
66828. 76575M D31 +9965 : b4,g5 c4,f5 d4,e6 e3,d6 f4,g4 h4,gxh3 gxh3,h6 h4 ...
81693. 93935M D33 +9965 : b4,g5 c4,f5 d4,e6 e3,d6 f4,g4 h4,gxh3 gxh3,h6 a3 ...
99800. 115279M D35 +9965 : b4,g5 c4,f5 d4,e6 e3,d6 f4,g4 h4,gxh3 gxh3,h6 a3, \
h5 h4,c6 c5,dxc5 dxc5,b5 cxb6,axb6 a4,c5 bxc5,bxc5
109912 127305M D35 16/16 : 1158.2 kN/s
0000011111 1112122222 2222222323 232310:::: :::::::::: :::::::::: ::::::::::
0447710436 5870922446 6778899090 918019:::: :::::::::: :::::::::: ::::::::::
1. d4? d5? 2.g4 wins (see the earlier d4d5 run)
1. d4? b5! draws (see D29)
1. c4 mate in 19 or 18
1. b4 mate in 18
Assuming no bugs of course.
... Johan
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.