Author: Robert Hyatt
Date: 18:55:06 05/20/05
Go up one level in this thread
On May 19, 2005 at 17:48:31, Vincent Diepeveen wrote: >On May 18, 2005 at 11:35:13, Robert Hyatt wrote: > >Is testing a testset automatically getting done deterministically or non >deterministically as compared to a single position that you set up? > >Vincent Depends on your definition of deterministic. If you use the "test" command inside crafty, then the hash tables are completely cleared between each test position. If you use the EPD facility, which uses a different set of commands to run the test set, then the hash table is not changed. However, since each position is searched for a fixed time interval, or to a fixed depth, the results are as repeatable as possible, since timing changes can always change a search result that is about to find a new best move just as time expires... > >>On May 18, 2005 at 00:37:13, Les Fernandez wrote: >> >>>Hi Bob, >>> >>>Was just wondering if you had the time to check on the answer to my post several >>>days ago. >>> >>>>That can happen, yes. I am not sure about how the EPD stuff is done, it might >>>>well clear the hash table anyway, I'll look... >>> >>>Thanks, >>> >>>Les >> >>Yes, but I forgot to respond, sorry. Turns out the EPD stuff does not force any >>hash table clearing. However, if you run the same test set repeatedly, this >>should still produce the same test results for almost all positions. There will >>be some variability in a very small number of positions, because the previous >>positions was searched slightly longer or slightly shorter the second run due to >>how time expires and terminates the search. Of course, if you cleared the hash >>table between positions, you would _still_ see this variability since the search >>could be almost ready to fail high and run out of time on one run, on the next >>run it does fail high before time expires...
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.