Author: leonid
Date: 17:36:43 03/16/00
Go up one level in this thread
On March 16, 2000 at 12:52:10, Robert Hyatt wrote: >On March 16, 2000 at 10:59:46, Georg v. Zimmermann wrote: > >>On March 16, 2000 at 10:45:39, Robert Hyatt wrote: >> >>>On March 16, 2000 at 09:42:43, William H Rogers wrote: >>> >>>>I am still having a problem with quiesce. Lets assume that after several plys >>>>my program ends up on e4. Do I just look at e4 to see if there are any captures, >>>>or do I search the whole board for captures etc.? >>>>I know that checks are included here too. >>>>Thanks in advance >>>>Bill >>> >>> >>>Some do it either way. I look at the whole board, find all the captures, >>>then weed out the ones that seem futile to search... >> >>Another question on that: >> >>Doesn't quiesce and extending on captures reach the same goal to a certain >>extend? What technique is better at what ? Do you do both ? >> >>Thanks. > > >In look at the 'search' as being broken into three unique and distinct >components: > >(1) the basic search. Here we look at _everything_ without discarding any >moves at all. We extend when we think the position warrants it (in check, >recaptures, etc.) IE we try to discover all the cute tactical stuff in this >part of the search, since what we find is pretty much error-free (except for >insufficient depth of course.) > >(2) the qsearch. Here we only want to make sure that there are no _simple_ >combinations that win or lose material. I don't personally want my q-search >to find wild tactical things, because it is _so_ error-prone when you mainly >search only captures. IE often the best response to a capture is not a >capture, but some other move with an even stronger threat. Since the q-search >is so error-prone, I try to limit what it sees, which limits these errors. > >(3) the evaluation. Which is basically blind to dynamic tactics. Since it >doesn't understand threat sequences and so forth, we try to make sure that >we don't give it such positions to evaluate, by letting the normal search and >the q-search carry us to a reasonably quiet position where the eval is correct. > >(3) is where almost _all_ programs make their mistakes. The eval says "white >is ahead by a piece, plus positional compensation" but it doesn't understand >that "black mates in 7 moves if we reach this position." Thanks, Hyatt! I found your explanation well done and useful for me as well. Leonid.
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.