Author: Nobuhiro Yoshimura
Date: 15:25:51 10/26/98
Go up one level in this thread
On October 26, 1998 at 09:34:33, Robert Hyatt wrote: >On October 26, 1998 at 01:36:59, Nobuhiro Yoshimura wrote: > >>On October 25, 1998 at 19:34:32, Robert Hyatt wrote: >> >>>On October 25, 1998 at 18:58:12, Nobuhiro Yoshimura wrote: >>> >>>>On October 22, 1998 at 23:20:10, Robert Hyatt wrote: >>>> >>>>>On October 22, 1998 at 23:04:31, Peter McKenzie wrote: >>>>> >>>>>>On October 22, 1998 at 21:38:29, Fawna Bergstrom wrote: >>>>>> >>>>>>>Well everyone has their opinions on this kind of question--here are a few of >>>>>>>mine. Let's go back to basics: >>>>>>> >>>>>>>Level I: You search full-width to a fixed depth (alpha-beta, iterative >>>>>>>deepening, etc. are all assumed, of course.) Here your evaluator includes both >>>>>>>material and positional factors. Move ordering is critical. First expand >>>>>>>"killer" moves, "interesting" moves and moves that yield a higher evaluation. >>>>>>> >>>>>>>Level II: If you like you can then search beyond that looking at "interesting" >>>>>>>moves such as captures, threats, checks, etc. Don't bother with threats unless >>>>>>>the threatened piece is hanging and/or more valuable than the threatening piece. >>>>>>> You should limit the depth of this second phase or you can skip it altogether >>>>>>>and go straight to level III--it's your call. In level II the evaluator >adjusts >>>>>>>for material only. Personally I wouldn't waste too many plies on Level II. >>>>>> >>>>>>I'm not 100% sure what you mean here, but if you start returning scores from the >>>>>> quiescence search that don't take into account changes in evaluation due to >>>>>>captures effecting pawn structure, and these scores can find their way into your >>>>>>PV, then you're likely to get killed positionally. >>>>>> >>>>> >>>>> >>>>> >>>>>there are things interesting to try here. IE if you only look at winning >>>>>captures, you can probably get away with a positional eval at the leaf node >>>>>that starts the qsearch, and then pass this value along to be modified as >>>>>pieces are captured. >>>>> >>>> >>>>May I ask a simple question ? I am a Japanese-Chess programmer so that >>>>I donot much about the chess programming. In the following postion: >>>> 0) In the q-search node >>>> 1) WTM >>>> 2) a white in danger >>>> >>>>Questions: >>>> 1) Do you generate esacaping moves for the white peice ? >>>> 2) Do you assume that a white piece can esacpe in the stand pat eval? >>>> 3) Is it better to make a "PASS" move ( without depth deductions) >>>> to check whether the black can really gain profit by capturing it? >>>> >>>> >>>>Nobuhiro >>>> >>>> >>> >>>I don't, no. I don't trust the q-search to do this sort of stuff. I only hope >>>I have extended enough already to avoid this problem. >>> >>>(2) yes as my eval doesn't consider pieces that are trapped/hung/etc except for >>>some special cases like a bishop at h2 with a pawn at g3 trapping it... >>> >>>(3) did this in the 1970's and it worked ok... ie "pass" then see what happens, >>>and *then* try to find a move that prevents that from happening... >>> >>>Don't do it today because of the extreme depths we see nowadays... >> >>In a position like: >> 0) in q-search node >> 1) WTM >> 2) white queen in danger >> 3) a black pawn in danger >> >>In such postion, do you only generate a capturing a black pawn? >>If so, the postion will be evaluted as a very bad position( losing queen). >>Are you saying that if we search deep enough, we donot get >>a bad side effect eventhough such position might occurs during the search? > >I'm only saying that after looking at all the games I have studied carefully, >I haven't seen a single game that was decided by some sort of failure in the >q-search. Maybe because I didn't search quite deeply enough, or because I >mis-evaluated something... but not because the q-search broke... > >but in the above case, I would simply not capture the black pawn. I would >first call Evaluate() to produce the stand-pat score, then try the capture. >At the next ply my opponent captures my queen and returns that score. I >decide that doing nothing is better than grabbing the pawn and losing my >queen, the idea being that instead of grabbing the pawn I can save the >queen. Which doesn't seem to cause problems except in rare cases where a >piece is really trapped and it doesn't realize this... Thank you! I understand now.
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.