Checks in QSearch

[David Weller]

Hi Dan,

No SEE at all
Every check imaginable
R=2

[Joachim Rang]

Hi Dan,

No SEE at all
Every check imaginable
R=2

With R=2 there is no need for checks in QS. Try R=3 wiht checks in QS and you may end better.

regards Joachim

[Tord Romstad]

I guess the extra ply gained by NOT doing checks, out weighs the advantage of possibly seeing MATE-IN-ONE eg., in the 1st ply of qsearch()

If checks at the first ply of qsearch cost you a full ply, I strongly suspect you have a buggy or inefficient implementation. It shouldn't be anywhere near that expensive. In Glaurung, checks in the qsearch rarely add more than about 20% to the node count, except in a few extremely wild tactical positions.

Tord

[David Weller]

Tord,

You are right, of course.

I mis-wrote. As I watched several of the games between the two versions, it seemed apparent that the one w/o/ checks was *occasionally* getting the extra ply - which doesnt mean it was actually getting a full ply deeper.

But what ever the trade was, it paid off, so it seemed.

so maybe its more like 1/3 or 1/4 ply deeper ??

Joachim,

The reason I added checks to qsearch() in the first place, was because of null-move[which I kinda love/hate, but thats another story:)]

But as usual, I might have only 'fixed' something incorrectly/accidentally then, by doing so - because now it doesnt seem as though I need them - so your point rings true, though before I might have been convinced otherwise.

I am considering the 'adaptive' null move scheme with r=2,3 but from someone who started out refusing to use null move at all, I am slow to move down this road....[I did try verified null move for a while with R=3 but gave up on it]

[Uri Blass]

The question is not number of nodes per ply but time per ply.

Fast check shows that in movei checks in the qsearch make movei 20% slower in nodes per second in the opening position.

Inspite of it I did not make the possibility to disable checks in the qsearch in movei personality and I do not believe it can help me.

I remember that at the time that I implemented checks in the qsearch I got no clear results if checks in the qsearch help it to play better and I decided that it is better because I know that my implementation of checks in the qsearch is not optimal and today I first generate all the legal moves and loop on them to look for checks.

Uri

[Uri Blass]

Note also that I had bugs in my implementation when it was not clear if it is better and it is another reason to believe that checks in the qsearch help me.

At that time I also used R=2 and I later found that R=3 is better for me.

Uri

[David Weller]

Hi all,

It seems I should read my own web page!

GES130 does NOT search checking moves!

I did play 25 40/5 games between GES130 and the same w/R=3 and checks in 1st ply of qsearch(). result was 12.5 : 12.5 which hints at least that the combinations are close.

I would have played more games, but ges crashed [oh no ...]

[David Weller]

Code: Select all

   Engine      Score                                GE
1: GES_130r3nl 17.0/32 =1=10==0=0==100===11110========1
2: GES_130r2nl 15.0/32 ????????????????????????????????

32 games played / Tournament finished
Name of the tournament: Arena tournament 10.18.04
Site/ Country: AMDATHLON, United States
Level: Blitz 1/1
Hardware: AMD Athlon(tm) Processor 1095 MHz
Operating system: Microsoft Windows XP Build 2600 Service Pack 1
PGN-File: F:\Program Files\Arena\Arena011.pgn
Website:
E-Mail Address:


1: R=3 plus checks in 1st ply of Q
2: R=2 no checks in Q