Author: James Swafford
Date: 13:33:50 11/15/99
Go up one level in this thread
On November 15, 1999 at 02:41:57, Peter McKenzie wrote:
>On November 14, 1999 at 22:03:09, Vincent Diepeveen wrote:
<snip>
>>
>>My nullmove implementation is basically allowing
>>doing a nullmove now if
>>
>> - king side to move is not in check
>> - one of both sides is not in pawn endgame
>> - last 2 consecutive moves were not nullmoves
>>
>>This last condition takes care you detect zugzwang too.
>
>This allowing of a double nullmove to detect zugzwang is a very elegant idea.
>I'm wondering, is anyone else (other than Vincent) using it? I plan to try it
>one day, but I have too many other things to do to LambChop first!
>
I just ran a short test with it. My first impression is that
I don't like it.
I only used the first 100 problems of the WAC suite.
The version of Galahad I did the test on has a very
crude eval (I'm preparing it for a GA experiment).
Anyway, from three different runs at 10 seconds per problem:
#1: r=2, control data
92 correct, 19 for mate
avg depth=9.061 (excluding problems solved for mate)
avg time to correct solution=.50 seconds
avg number of nodes (until time expired)=1,540,574
avg nps (excluding mates)=188,826
#2: r=3 near the root, r=2 when depth <= 6
92 correct, 19 for mate
avg depth=9.197
avg time=.50 seconds
avg number of nodes=1,499,775
avg nps=183,790
#3: same, with "double null move"
91 correct, 19 for mate
avg depth=8.074
avg time=.45 seconds
avg number of nodes=1,489,178
avg nps=182,306
I think that run 2 indicated an improvement. It utilized
Ernst's idea that I read about in another post. Although
marginally slower, it searched marginally deeper. :-)
Run three was disappointing. One problem was lost,
and depth dropped a full ply. The only plus was that
time to solution (avg) dropped to .45, but this can be
misleading. It calculated for 91 correct problems, not
the 92 that run 2 had. If I dug down into my log files,
I'd probably find one problem in runs 1 & 2 that was barely
solved in time.
I know the intent here is to avoid zug problems, and it
probably does help that, but at what cost? Is it worth
it? Maybe if the 'threshold' from run 2 was used until
the endgame, and *then* adding Vincent's "double null move,"
were zug is much more likely, and tactical depth isn't
as much of a problem.
Of cource, Vincent is welcome to tell me I don't know
what I'm talking about. :-)
--
James
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.