Author: Andrew Williams
Date: 03:58:44 04/24/01
Go up one level in this thread
On April 24, 2001 at 04:08:12, Tony Werten wrote: >On April 23, 2001 at 13:37:19, Andrew Williams wrote: > >>On April 23, 2001 at 11:13:15, Ulrich Tuerke wrote: >> >>>On April 23, 2001 at 10:58:27, Dieter Buerssner wrote: >>> >>>>On April 23, 2001 at 10:11:00, Ulrich Tuerke wrote: >>>> >>>>>On April 23, 2001 at 09:48:40, Dann Corbit wrote: >>>>[...] >>>>>> >>>>>>Just like NULL MOVE or any selective element for searching, it will cause some >>>>>>problems to be solved more slowly. However, I suspect it will cause *most* >>>>>>problems to be solved faster. That is the way it works with every good idea. >>>>> >>>>>Completely agreed. The above mentioned risk arises essentially if you have >>>>>chosen your cut threshold (too) close to alpha. >>>> >>>>There can also be some other "risks", depending on what the qsearch does. >>>>E.g. When you don't allow stand-pats, where the side to move is in check. >>>>When you cut such nodes, depending on material gain and alpha, you can miss such >>>>situations. >>>> >>>>I try to be careful to do the pruning in qsearch in certain situations. When you >>>>capture the last pawn of the opponent, you can reach a draw score. An example >>>>(similar to what we discussed recently). KNNP vs. K will have an high material >>>>advantage. If you grap the P, the material gain is much less than the gain in >>>>score. Also, when reaching pawn endgames, I try to be careful. >>> >>>Yes. I made similar experience. Quite generally, futility pruning in pawn >>>endings implies an unreasonable high risk. I think that this holds for the full >>>search as well as for the q-search. >>> >>>Neither do I cut when the colour to move is in check. >>> >>>> >>>>I think, this pruning idea saves less, than one could think. When you prune the >>>>node, all work, that would be needed for this node is a call to qsearch, and a >>>>call to eval (which will fail high). >>> >>>In comet, it does pay out. I observed that I win O(20%) by applying this kind of >>>q-search pruning, depending on the threshold chosen and position. >>> >>>> >>>>For the threshold, I use something based on largest positional advantage >>>>(dependant on the side). Of course, this won't give many cutoffs with Vincent's >>>>20 pawns :-) >>> >>>I use about a half pawn, which is perhaps a bit too aggressive. >>> >> >>Interesting. I use the highest positional score so far in the search, >>with a *minimum* of 2 pawns. > >You don't have to do that, since you made a call to evaluate in the start of >qsearch. Your threshold only has to be the largest gain you got from 1 move. > >Tony > Yeah. I'm coming round to the view that I'm definitely being over-cautious. Andrew
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.