Author: Gerd Isenberg
Date: 10:52:13 03/08/06
Go up one level in this thread
On March 08, 2006 at 01:24:31, Gerd Isenberg wrote:
><snip>
>>>>>your eval or other odd things like that).
>>>>
>>>>** I do use floating point for the evaluation. This is a relic of something
>>>>** that can be pulled out of the program if it is a really bad thing. Bob
>>>>** has said it is due to floating points always being off.
>>>>
>>>
>>>Floats just for eval or also all bounds and scores backed up to the root?
>>>SSE floats or doubles under w64 are quite efficient, for instance you (or your
>>>compiler) can work with vectors of four floats per instruction.
>>>Is your nullwindow {alfa, alfa+1.0} or something like {alfa, alfa+1.0e-10}?
>>
>>They're all doubles and for everything that would normally be an int.
>>
>>This is handled with a typedef and could fairly easily be an int with
>>some additional ifdefd code for %d as opposed to %f.
>>
>>My null window is always -alpha-1,-alpha.
>>
>
>while 1 is the smallest int greater zero, i wonder with float or doubles whether
>there is no "smaller" null window dependent on a possible fractional part of
>your evaluation. Did you try -alpha-epsilon,-alpha for null windows with epsilon
>far less one?
In PVS with null windows - if no cut occurs - you don't improve alfa.
This is obviously not the case if your alfa is let say 10.0 and beta 11.0 and
your score is 10.1 or 10.9999.
>
>>If you think double is severely affecting reproducibility or putting
>>bugs in that could cause a performance-hurting issue, I can make it
>>int.
>>
>>Stuart
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.