Author: Robert Hyatt
Date: 20:37:11 08/10/04
Go up one level in this thread
On August 10, 2004 at 21:52:28, Stuart Cracraft wrote: >On August 10, 2004 at 10:35:29, Anthony Cozzie wrote: > >>On August 10, 2004 at 09:35:19, Stuart Cracraft wrote: >> >>>Enclosed please find SEE results for SEE in quiescence >>>(SEE >= 0 to search a capture), SEE in all move sorting >>>(both quiescence and main search - move score is see() value >>>plus a centrality value plus history heuristic for captures, >>>for non-captures only centrality value plus history heuristic) >>>and lastly SEE for both of the above. >>> >>>No SEE: >>> >>>$ ./qaNbatch 30 300 >>>HERALD ga nosee.log 30 300 >>>**** 9.97/35.22 85% 255/300 7187.14 1756630528 5855435/24/244413 0/0/102630520/0 >>>/0/0 >>> >>>SEE in Quiescence >>> >>>HERALD ga -DSEEQUIESCE seequiesce.log 30 300 >>>**** 10.13/31.83 85% 257/300 7147.18 1572046592 5240155/24/219953 0/0/51211448/0 >>>/0/0 >>> >>>SEE in move sorting >>> >>>HERALD ga -DSEESORT seesort.log 30 300 >>>**** 9.53/35.23 82% 246/300 7267.69 1423642368 4745475/24/195887 0/0/85531800/0/ >>>0/0 >>> >>>SEE in both quiescence and move sorting >>>HERALD ga -DSEEQUIESCE -DSEESORT seeboth.log 30 300 >>>**** 9.81/32.28 84% 253/300 7207.02 1331241728 4437472/24/184715 0/0/43808052/0/ >>>0/0 >>> >>>The bottom line result is that the average depth reached was 1/5 ply deeper >>>using just SEE in quiescence compared with no see and this was good enough for a >>>slightly higher (<1%) test score result. >>> >>>Other uses of see in move sorting and for both did not do as well as for just >>>limiting the capture search. >> >>Move ordering in general is more important as depth increases. If a move >>ordering change decreases nodes by 5% at depth 6, that means an entire ply at >>depth 14 ;) >> >>Plus, if you have only PST eval you should be getting 3-4M nps, so SEE probably >>slows you down a _lot_. Once you get a more complex engine you will find that >>SEE move ordering is critical. >> >>anthony > >I don't know what kind of PC's you guys have but I don't think my >implementation is *that* bad to get only 1/11th of the speed you expect >for PST. > >Unless I'm missing something really major in the search implementation. On a single-cpu 2.8ghz pentium 4, Crafty gets around 1M nps for reference. On my 750 mhz PIII laptop, it gets around .5M nodes per second...
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.