Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: SEE results

Author: Robert Hyatt

Date: 15:50:14 08/10/04

Go up one level in this thread


On August 10, 2004 at 17:36:53, Uri Blass wrote:

>On August 10, 2004 at 14:47:02, Robert Hyatt wrote:
>
>>On August 10, 2004 at 13:42:24, Eric Oldre wrote:
>>
>>>On August 10, 2004 at 12:52:28, Robert Hyatt wrote:
>>>
>>>>On August 10, 2004 at 10:59:29, Tord Romstad wrote:
>>>>
>>>>>On August 10, 2004 at 10:35:29, Anthony Cozzie wrote:
>>>>>
>>>>>>Plus, if you have only PST eval you should be getting 3-4M nps, so SEE probably
>>>>>>slows you down a _lot_.
>>>>>
>>>>>Yet another proof of how bad my programming skills are.  With PST eval and
>>>>>nothing
>>>>>else, I get about 800,000 nps (on a PIV 2.4 GHz).  Adding SEE slowed me down to
>>>>>around 750,000 nps.
>>>>>
>>>>>Tord
>>>>
>>>>
>>>>SEE slowed me down by 10%.  Move ordering improved to speed me up 10%.  Washed
>>>>out to no advantage, _until_ I added the stuff about tossing out captures that
>>>>can't bring the score back to within the alpha/beta window.  That made me about
>>>>2x faster overall, which is a _big_ gain.  Stuart isn't getting all the
>>>>advantages of SEE yet, and, in fact, It isn't necessary to use SEE on _every_
>>>>capture.  IE for PxQ there is no need to see if that is +9 or +8.  Either is
>>>>good enough and there MVV/LVA is faster.
>>>
>>>
>>>
>>>Robert,
>>>I haven't done any extensive testing to see if SEE pruning would improve the
>>>strength of my engine. but i was worried that pruning moves in qsearch based on
>>>a weak see score would eliminate moves where there would be additional
>>>discovered attacks.
>>>
>>>For now I settled on the following. Over time i'll be able to tune the 2 const
>>>values of 75 and 200 to more appropriate levels. I'm guessing evenutally they
>>>will be lowered, but for now i'm trying to be "safer".
>>>
>>>
>>>//get the see value of the move
>>>move_estval = movestack->estscore[movestack->orderedmove[imove]];
>>>
>>>//get the captured piece
>>>chesspiece captured_piece = MoveTaken(move);
>>>
>>>//find value of the captured piece
>>>int captured_piece_val = evalterms.material_vals[captured_piece];
>>>
>>>bool can_gain_mater = true;
>>>//init_score is the static eval before moving
>>>if(captured_piece&&((captured_piece_val+init_score+75)<alpha)){can_gain_mater=false;}
>>>if(captured_piece&&((move_estval+init_score+200)<alpha)){can_gain_mater=false;}
>>>if(can_gain_mater){should_do_move=true;}
>>>
>>>In your opinion, is there any truth to my concerns about not discovering attacks
>>>because of a see score that couldn't reach alpha?
>>>
>>>Eric
>>
>>
>>I consider SEE scores to be simply estimates, since they do not worry about
>>pinned pieces and the like.  But then the entire q-search is only an estimate
>>since it doesn't consider anything but captures (in Crafty).  There are enough
>>errors in only looking at captures to make your hair stand on end.  Compounding
>>this with SEE pruning doesn't change much at all.
>>
>>Best test is to try it.  IE don't prune losing captures, and run a big test set.
>> Then prune them and run it again.  You will probably get almost a ply deeper by
>>pruning, and your scores should go up, not down as a result (at least when I
>>tested this it did.  I've been doing this since the late 70's, with reasonable
>>results.  Others such as Bruce (Ferret) Stanback (Zarkov) and so forth have had
>>the same results).
>>
>>The right thing to do is shrink the q-search, where the errors are, and make
>>your normal search bigger where there is none of this highly selective pruning
>>(ie looking at nothing but captures is _really_ selective) and its associated
>>errors.
>
>I think that the right thing to do is to be more selective in the last plies.
>
>It does not mean to do the qsearch smaller.
>You can start with more selective search than normal search but less selective
>search than your qsearch in the first plies by including checks and captures and
>only later to consider only good captures(I include queen promotions in the word
>capture).
>
>Uri

I don't disagree.  Cray Blitz did that.  It had a normal full-width part of the
search, then a "tactical phase" that included captures, checks and some moves
with significant threats, then only captures/checks, and finally tapering off to
only captures.

I may try that again one day as I did play with it in early versions of Crafty
but decided to "keep it simple" for a while...





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.