Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: q-search question

Author: Landon Rabern

Date: 11:48:13 07/20/00

Go up one level in this thread


On July 20, 2000 at 14:35:56, Bas Hamstra wrote:

>On July 20, 2000 at 12:24:49, Landon Rabern wrote:
>
>>On July 20, 2000 at 08:15:17, Bas Hamstra wrote:
>>
>>>On July 19, 2000 at 14:51:17, Landon Rabern wrote:
>>>
>>>>On July 19, 2000 at 14:08:23, Andrew Dados wrote:
>>>>
>>>>>On July 19, 2000 at 14:00:16, Landon Rabern wrote:
>>>>>
>>>>>>On July 19, 2000 at 13:22:47, Andrew Dados wrote:
>>>>>>
>>>>>>>On July 19, 2000 at 12:03:10, Landon Rabern wrote:
>>>>>>>
>>>>>>>>I have been discarding all captures where attackervalue>DefenderValue in my
>>>>>>>>q-search which speeds it up significantly, but I know that it is throwing away
>>>>>>>>some captures that are good.  So I implemented a SEE function.  The SEE returns
>>>>>>>>the correct value on tests I have run.  When I put this into my program so that
>>>>>>>>if (attackervalue>DefenderValue)&&(SEE>=0) I keep the move as well, I got worse
>>>>>>>>results on the WAC test suite.  Before I put the see in I got 270/300 at 60
>>>>>>>>seconds per move and after I got 257/300 at 60 seconds per move.
>>>>>>>>
>>>>>>>>Is it just that there are no capture sequences in this test that need the extra
>>>>>>>>captures, or is there something wrong with my SEE function?
>>>>>>>>
>>>>>>>>Thanks for any help,
>>>>>>>>
>>>>>>>>
>>>>>>>>Landon W. Rabern
>>>>>>>
>>>>>>>It would be of help if you post some relevant positions. I may just guess now
>>>>>>>that if you do check detection in qsearch you may find some mating combinations
>>>>>>>with 'losing captures', when recapturing piece is overloaded simply, so in next
>>>>>>>move capture is mate. (Or you may do some non-capturing,checking moves in
>>>>>>>qsearch which complicates matters still). With SEE you miss those. Question is
>>>>>>>if average speedup of SEE in non-tactical positions offsets those few missed by
>>>>>>>using SEE....
>>>>>>>
>>>>>>>-Andrew-
>>>>>>
>>>>>>I do not do checks in q-search.  The problem is that I should be missing fewer
>>>>>>tactical positions with SEE, than with just throwing all captures out where
>>>>>>attackerValue>defenderValue.
>>>>>>
>>>>>>Landon
>>>>>
>>>>>So maybe you are much faster then SEE with throwing away all captures where
>>>>>attackerValue>defenderValue... Single out those positions and compare PVs and
>>>>>nodes produced with both versions of your qsearch ply by ply. Then maybe you/we
>>>>>can learn what's going on. Do you reach extra ply or does your program find
>>>>>those moves one/few plys sooner?
>>>>>
>>>>>-Andrew-
>>>>
>>>>I rememeber reaching an extra ply in many cases, I can not take a look at the
>>>>PV's until late tonight since I am at work now and will be working late to get
>>>>the product out on time.
>>>>
>>>>Landon
>>>
>>>I cannot imagine throwing A>D captures out of the qsearch works. Even if it
>>>solves many testpositions, it cannot be accurate in games.
>>Would you mind trying it out to see if you get similar results?  I throw
>>everything out where A>D except I keep all captures where the captureer is a
>>king.
>
>I tried it. It did bad. It solves less positions and searches less deep on
>average. I also compared with no pruning at all and SEE pruning. The latter two
>solve an equal amount op positions and search equally deep.
>
>
>SEE pruning:
>
>Positions:			300
>Right:		 		200
>Wrong:		 		100
>Right perc:			66.45
>Nodes:				4162440
>QNodes:				3877382
>Total:				8039822
>Average Depth:		        5.29
>
>
>No pruning:
>
>Positions:			300
>Right:		 		201
>Wrong:		 		199
>Right perc:			66.78
>Nodes:				4166168
>QNodes:				10198475
>Total:				14364643
>Average Depth:		        5.30
>
>
>Your A>D except for king pruning:
>
>Positions:			300
>Right:		 		198
>Wrong:		 		102
>Right perc:			65.78
>Nodes:				4162033
>QNodes:				3097453
>Total:				7259486
>Average Depth:		        5.26
>
>All three runned in the same (short) amount of time. Note that for this older
>program of mine SEE pruning breaks even. It cuts a lot of nodes, but still
>searches to the same depth.
>
>Your A>D pruning does slightly worse. But I think this "slightly" is misleading.
>This system will kill you in real games.
>
>
>>>Furhter I think you should compare SEE pruning with no pruning at all. And then
>>>you should see that SEE goes deeper on average. Or else your SEE has a bug OR
>>>your SEE is too slow.
>>
>>With SEE it does go deeper on average than with nothing.  And my SEE function is
>>very fast.
>>
>>>
>>>Good way to debug your SEE is to compare the SEE result with the QSearch result.
>>>If this doesn't work, make a new special qsearch, that simulates SEE behaviour.
>>>It doesn't matter if it is slow, as long as it works accurate. Comparing the two
>>>will show you if there are bugs.
>>
>>I set up a bunch of mock positions that had a bunch of capture/recapture and ran
>>through them on paper and then with the SEE, same result, so I think it is
>>correct.
>>
>>Last night I ran WAC with no pruning in the qsearch and I got 253/300 compared
>>to with SEE 257/300 and with A>D plus king caps 270/300.  The I ran on WCSAC and
>>got 745/1002 with SEE and 752/1002 with A>D +king caps.
>>
>>
>>If anybody has ever tried just cutting with A>D + king caps let me know.  I know
>>that it should be doing worse since it would leave a hung pawn that could be
>>captured by a queen at the end of the search.
>>
>>Thanks,
>>
>>Landon

Thanks, I think I know what the problem is now.  You ran that at a very short
time control, right?  I ran at 60 seconds per move getting 9-14 plys deep, so I
think I outsearched the main need for the q-search, I searched past the
capture/recaptures in my main search and with the SEE it was just slowing this
down a little causing me to miss more positions.  I'll bet if I run it with no
qsearch, I will get even better test results, but thats not good, so I will be
switching to using SEE.


Thanks again for trying that out,

Landon



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.