Author: Vincent Diepeveen
Date: 11:14:49 03/27/01
Go up one level in this thread
On March 26, 2001 at 17:36:01, Robert Hyatt wrote: >On March 25, 2001 at 12:21:56, Vincent Diepeveen wrote: > >>On March 25, 2001 at 09:37:09, Robert Hyatt wrote: >> >>>On March 25, 2001 at 03:51:02, Pham Minh Tri wrote: >>> >>>>Hi, >>>>I am wondering why we do not search checking moves (moves check opponent king >>>>but do not capture) in qsearch function. I think they are not ?quiescence? and >>>>should be searched as capture moves. Does anyone try them? >>>>Thank very much for any explanation. >>>>Pham >>> >>> >>>I have two answers: >>> >>>1. I did q-search checks in Cray Blitz. I also did them in early versions >>>of crafty as you can see in the comments in main.c... >>> >>>2. I removed them a _long_ while back in Crafty. >>> >>>Here is my analysis: If you do q-search checks, you will find some tactical >>>lines significantly quicker, obviously. But since you are searching much >>>deeper along lines with lots of checks, overall you have to search less deeply >>>along non-checking lines, since you can only search a fixed number of nodes in >>>a fixed unit of time. >> >>>You have to make the following choice: either (a) find tactical shots that >>>depends on checks quicker, but give up some positional skills as positional >> >>As doing checks in qsearch finds tactics 4 ply sooner for me, sometimes >>even zillions of plies as i try unlimited number of checks in many cases >>in qsearch, for me it's quite easy to go for a here without ever considering >>b. Idem for many commercial programs. > >I wouldn't begin to argue here either way. But if you find tactics 4 plies >sooner, how much depth do you lose overall because of those checks? Is it >worth the lost depth to see tactics sooner? Many positions i win depth, i don't lose depth there. The better score returned from the qsearch gives better values in the pvs mainsearch, so cutting down big trees. Obviously many positions it DOES need more nodes. But what i also do in qsearch is many captures which you don't even try in crafty, so the bigger overhead is basically those captures NOT che checks. >If I lose enough games due to this, I will certainly implement it once again. About all blitz losses against Tiger and other commercial programs are because of this. At least 80% of all games you lose in blitz on icc is because you don't do checks in qsearch. But i have to agree with you that if you can't get it efficient implemented in qsearch, that it's a big hell to use checks in qsearch! Also your SEE routine needs revision then, as it's useless to try checks in qsearch but not try to optimize scorees with some easy captures. >But I haven't, so far, and I like the 14 ply depths in normal middlegame >positions. That covers up a lot of potential positional mistakes.. That compares to like 9 ply search of DIEP, from tactical reasons seen when it has to do with check. > > > >> >>>moves will be searched less deeply; (b) find tactical shots slower, but then >>>search deeper in positional cases and find positional threats your opponent >>>won't be able to see because the checks prevent him from getting deep enough >>>to see it. >>> >>>I like the non-check approach because (1) I can search deeper overall, and in >>>non-tactical positions this pays off well; (2) the q-search code is much >>>simpler, which makes it even faster since so much of the total time is spent >>>there; >>> >>>Of course, everyone should try it themselves to make sure they get the same >>>results I get. I do know of several programs that use this approach with >>>great success. Note that I don't particularly think much of the idea of >>>searching checks very deeply but overlooking all of the _other_ kinds of >>>tactical moves you might include in the q-search. Pins. Attacking overloaded >>>pieces. etc... >>> >>>By keeping the q-search simple, the normal search handles more of the nodes, >>>and since that part of the search includes _everything_ you might be tactically >>>stronger by not overlooking something.
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.