Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: q search nps question

Author: Robert Hyatt

Date: 20:15:53 01/07/00

Go up one level in this thread


On January 07, 2000 at 18:45:14, Landon Rabern wrote:

>On January 07, 2000 at 12:25:46, Robert Hyatt wrote:
>
>>On January 07, 2000 at 03:56:37, Landon Rabern wrote:
>>
>>>I jsut added a q-search to my program and it dropped from 180,000 nps to
>>>90,000 nps.  I count leaf nodes when counting my nodes, so I though it might be
>>>that the q-search has no leaf nodes, so I was not getting these free no-work
>>>nodes.  This wasn't it though, because I tried incrementing the nope counter
>>>where a leaf node would have been had it been a regular search, but this only
>>>improved it slightly.  I am pretty sure that it is actually running a lot
>>>slower.  Is it supposed to do this?
>>>
>>>Thanks,
>>>
>>>Landon W. Rabern
>>
>>
>>The issue is how do you generate captures.  If you generate all moves, and
>>then extract the captures for the q-search, your nps must drop, because of
>>all the extra work you do but never take advantage of.  The q-search is
>>more selective, meaning that for each node you reach in the q-search, you do
>>the usual amount of work in the alpha/beta search, and so forth, but you only
>>look at a couple of branches.  That drops the NPS significantly.
>
>
>I generate only captures and only use the captures that are "good" meaning that
>the piece capturing is worth the same or less than the one captured.  Do you
>think adding the captured king method of testing for check in the q-search will
>about even things out?
>
>Landon W. Rabern


No.  At internal nodes, you do all the alpha/beta work, the procedure call
to search, etc, and ammortize that over a bunch of moves. In the q-search,
you do the same sort of set-up work, but ammortize it over a couple of moves
at most.  NPS must go down....



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.