Author: Alex Szabo
Date: 10:17:56 12/06/03
Go up one level in this thread
On December 06, 2003 at 12:38:25, David Dahlem wrote: >On December 06, 2003 at 12:31:41, Alex Szabo wrote: > >>On December 06, 2003 at 11:22:15, Alex Szabo wrote: >> >>>My self-play experiment between crafty versions 19.6 and 17.14 resulted in the >>>19.6 version being stronger by 11 (plus or minus 2) rating points. >>> >>>The experiment was for 20,000 games at a 20+1 time control. It was run on a >>>2x2GHz pentium4 linux system with xboard 4.2.6. I can include other details if >>>there is interest. >> >>Here are some more details: >>(1) 4 matches were run at the same time >>(2) pondering was on (ie normally 8 craftys were competing for CPU time) >>(3) hyperthreading was disabled (FYI, when enabled there are 4 processors and >>5-10% better nodes/sec on this type of computation) >>(4) use of 3 & 4 piece tablebases was enabled and cache set to 32M >>(5) learning was disabled >>(6) there was no opening book; games started from 10,000 different opening >>positions compiled from a database of grandmaster games >>(7) each unique starting position was played twice with players reversed > >How were you able to determine that each starting position was unique? I started from gm2600.pgn (from ftp://ftp.cis.uab.edu/pub/hyatt/common) and truncated the games at the earliest point that the move sequences could all be unique (or 200 bytes max ascii move sequence length) and deleted any duplicates. I got 20,326 unique move sequences from which the 10,000 are a sample. It is very likely that some of these positions are the same because of move transposition. My statement above in detail (6) is not completely accurate; perhaps it should read "...10,000 mostly different..."!? > >Regards >Dave > >>(8) all other crafty settings were default >>(9) used 17.14 compiled with icc (-O3 -march=pentium4) which gets 849,011 n/s on >>bench command >>(10) used 19.6 compiled with icc (-O3 -march=pentium4) which gets 796,900 n/s on >>bench command
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.