Author: Robert Hyatt
Date: 12:35:17 10/06/99
Go up one level in this thread
On October 06, 1999 at 10:23:21, Eelco de Groot wrote: > >On October 06, 1999 at 10:01:14, Robert Hyatt wrote: > >>On October 06, 1999 at 04:21:31, Eelco de Groot wrote: >> >>>Robert, Eugene, what does prevent Crafty of adopting the better search of Cray >>>Blitz? I understand there are big architectural differences but what do you >>>think would be the main factor? >>> >>>Thanks for any comments, >>> >>>Eelco >>> >>> >> >>Not a thing. IE I have been experimenting with singular extensions (used in >>Cray Blitz) off an on. I just haven't found anything that I particularly like >>yet... But you can bet that nothing inside cray blitz will be lost, except >>for the vectorizabile stuff that will perform very poorly on a PC. But other >>search ideas used there will eventually be tried in Crafty... >> >> > >Robert , you do an awful lot of your testing on ICC at Blitz-timecontrols. >Maybe singular extensions would do better at longer timecontrols? I remember >having read that somewhere, that the results were better with deeper searches. > >Regards,Eelco Most testing is at blitz on ICC. But I pay much less attention to the results. IE Crafty's rating has been much more stable in standard than in blitz, and I have been paying much more attention to how it does at longer time controls. I do notice things at blitz, but I make sure that before I change something, that it doesn't make it worse at longer time controls. For things like SE, I don't use ICC to test that. I use test suites. IE, if I can get 300 WAC positions in 60 secs per move without, I would hope that I can get 300 right in less than 60 seconds per move with SE. That is how I generally test. And then I have some non-tactical positions (ie the kopec test suite) that I run to be sure that in non-tactical positions I am not signficantly slower than before... For testing things like extensions, ICC isn't the right place. ICC is good for overall evaluation tuning experiments, but a single extension might not affect but 1 in 10 games, so that would be non-optimal test results if ICC was the primary evaluation tool...
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.