Author: Robert Hyatt
Date: 05:44:14 09/17/98
Go up one level in this thread
On September 17, 1998 at 08:24:43, Pat King wrote: > >On September 16, 1998 at 21:42:45, Robert Hyatt wrote: > >>On September 16, 1998 at 17:16:20, John Coffey wrote: >> >>>On September 15, 1998 at 21:11:54, Robert Hyatt wrote: >[snip] >>correct... my "basic search" is an attempt to search active positions to the >>point they are almost "quiet" or totally "quiet" by doing the various extensions >>I do. Then I tack on a capture-only search to clean up the last bit of >>uncertainty when possible... so that the static evaluator doesn't have to >>understand hung pieces, forked pieces, whether the king is in check or not, >>and so forth... >I do the same thing, but I want to be sure I understand something... >For the static evaluator to be blissfully unaware of checks, I have a >check extension in my qsearch as well as my search. Further, I allow the >checked side to evaluate all moves, not just captures. And as a result, >I sometimes get the runaway evaluation that John described at the start >of this thread. Isn't this approach neccessary to simplify the evaluator >as you have described, and if so, then isn't an absolute upper limit to >the search depth required? > >Pat I simply ignore checks in the qsearch. I followed them carefully in Cray Blitz, going to the trouble to note when following checks is useful (ie if all of the moves by a side are checks in the q-search you can find a forced mate, but if there is one non-checking move, then you can't find forced mates below that node because if he is not in check, your opponent can simply stand pat and not ever get mated.) I depend on my normal search extensions to "ride out" the checks. Then, once I get to the q-search, I try to catch hanging pieces/overloaded pieces, and so forth, but only with respect to captures. Here's why: the qsearch is *full* of errors. There are zillions of positions where the best response to a capture is *not* a capture. Perhaps a rook move pinning the last capturing piece so you can win it outright, rather than just winning a pawn, for example. If the q-search is full of errors, and it definitely is, I don't want to rely on *that* to win material, because it is going to be very inaccurate. So, I choose to spend more time in the search above the q-search, where I look at all moves anyway, and where there are far fewer errors to contend with. IE I want my normal search to win material, because it is probably right, I don't want my q-search to try to win material because it is not trustworthy...
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.