Author: Vincent Diepeveen
Date: 11:05:14 04/26/04
Go up one level in this thread
On April 26, 2004 at 13:51:03, Robert Hyatt wrote:
>On April 26, 2004 at 11:50:18, Vincent Diepeveen wrote:
>
>>On April 26, 2004 at 06:31:58, Tord Romstad wrote:
>>
>>I wonder how you came to your conclusion.
>>
>>I did more experiments and concluded that always nullmoving saves most time.
>>Double nullmove is just performing slightly worse (less than a percent in search
>>time) and not missing zugzwangs, so i use it for that reason.
>
>
>This is one of those rare occasions where I agree with Vincent.
>
>Several years ago Bruce Moreland and I were talking, about this very subject. I
>was explaining how I did it in Cray Blitz (and also in Crafty). I classified a
>node as "ALL", "CUT" or "PV" according to Marsland/Schaeffer terminology, and if
>the node was classified as PV I avoided doing the null-search at all.
>
>Bruce did some testing and doing this restriction actually slowed him down a
>bit. So we both set about testing this stuff extensively and we came to the
>conclusion that null-move is best used _everywhere_ because we don't have
>perfect ordering and node classification is not that accurate as a result. The
>only reason for avoiding a null is:
>
>(1) you are in some sort of zugzwang position where a null-move will fail high
>for the wrong reason and wreck the search. Classic examples here are positions
>with very few pieces. IE pawns vs a knight where the knight can be zugged.
>Most require some minimal amount of material on the board to avoid this problem.
>
>(2) there is a tactical issue that is hidden with the R reduction. IE the
>classic position with white pawns at f2, g3 and h2, black queen at h3 and black
>pawn or bishop at f3, threatening mate on the move. If the R reduction prevents
>you from seeing the mate, you can have problems.
>
>(3) The hash table proves that the null-move search will not fail high, meaning
>that the search will be wasted effort.
>
>(4) Obvious positions such as when the side on move is in check. Not moving
>can't fail high here as the king is lost.
>
>(5) I don't allow two consecutive nulls. It is a potentially cute way of
>eliminating zugzwang problems, but it is only good for that, and it is not free
>in positions where no zugzwang is possible. I choose to not deal with it
>although I have this on my "to do" list to test with (say) pawn-only endings.
Not allowing 2 consecutive nulls will for sure not find zugzwang problems.
Not allowing 3 consecutive nulls will however find many zugzwang problems. As
pointed out hashtable could do difficult, but that very seldom happens.
Double nullmove does allow 2 consecutive nulls but not a 3d. So for a correct
implementation you just need 1 'if then else' to also not allow null in a
transposition table cutoff. This is very trivial to do.
>In short, try null everywhere. But, as always, test, test, test. Theory says
>that doing null on a PV node is a waste. Practice says that not all PV nodes
>are really PV nodes after the search has started... Go with practice. :)
I'm using PVS in its pure implementation, so i start with window [-inf;inf]
i also disallow nullmove when beta == inf.
That's trivial stuff however.
All other 'pv nodes' doing a nullmove is very good for different reasons. Even
when it gets a nullmove. double nullmove has a bit of a IID effect.
>>Your results give IMHO only the indication that your qsearch needs to be fixed
>>and/or that you did your experiments at a too fast time control where tactics
>>dominate.
>>
>>>Last week, there was an interesting discussion on this forum about which
>>>criterions should be used to decide whether or not to do a null move search.
>>>In an attempt to find out what works best for my engine, I played a self-play
>>>tournament between four slightly different versions of my engine. The
>>>following pseudo-code explains the null-move criterions used by each of
>>>the four versions:
>>>
>>>/* Gothmog A: */
>>>int ok_to_do_nullmove() {
>>> if(in_check(SideToMove)) return 0;
>>> if(static_eval - value_of_biggest_hanging_piece(SideToMove) >= beta)
>>> return 1;
>>> return 0;
>>>}
>>>
>>>/* Gothmog B: */
>>>int ok_to_do_nullmove() {
>>> if(in_check(SideToMove)) return 0;
>>> if(static_eval - value_of_biggest_hanging_piece(SideToMove)
>>> + value_of_biggest_hanging_piece(Opponent) >= beta)
>>> return 1;
>>> return 0;
>>>}
>>>
>>>/* Gothmog C: */
>>>int ok_to_do_nullmove() {
>>> if(in_check(SideToMove)) return 0;
>>> else return 1;
>>>}
>>>
>>>Gothmog D is similar to Gothmog C, except that when the remaining depth
>>>is 4 plies or more, a null move search with the depth reduced to zero
>>>is done before the real null move search, and the real null move search
>>>is avoided if the zero-depth search fails low.
>>>
>>>The criterion in Gothmog A is the one I have always used in the public
>>>versions of Gothmog. Our discussion last week seemed to indicate that
>>>most others use the simpler criterion in Gothmog C.
>>>
>>>In my tournament, I all four versions face each other 50 times. The
>>>results:
>>>
>>> Gothmog A Gothmog B Gothmog C Gothmog D Sum
>>>Gothmog A XXXX 26.5 31.0 29.5 87.0/150
>>>Gothmog B 23.5 XXXX 30.5 26.0 80.0/150
>>>Gothmog C 19.0 19.5 XXXX 31.5 70.0/150
>>>Gothmog D 20.5 24.0 18.5 XXXX 63.0/150
>>>
>>>It is interesting that the version with the most restrictive criterion
>>>achieves the highest score. This confirms my growing suspicion that
>>>recursive null move pruning doesn't work nearly as well as most people
>>>believe, and that it might be a good idea to make an effort to reduce
>>>its use as much as possible. I will try to experiment with even more
>>>restrictive criterions (perhaps static_eval-biggest_hanging >= beta+margin)
>>>and see whether that improves the strength further.
>>>
>>>The number of games is of course still not sufficiently big to make any
>>>definitive conclusions. If I find the time, I will add more games and
>>>more versions to the tournament.
>>>
>>>Tord
This page took 0.02 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.