Author: Don Dailey
Date: 12:09:13 12/23/97
Go up one level in this thread
On December 23, 1997 at 13:27:22, Stuart Cracraft wrote: > >What are the most useful kinds of extensions? > >I've heard of five: > > singular - if one move is clearly better than all other moves > > promotion threat - if a pawn moves to the 7th rank > > recapture - a recapture restoring previous material difference > > check - if a move is a check > >What other extensions are there that have proven useful? > >Stuart Cracraft Hi Stuart, Lot's of things are possible. Unfortunately I have had little luck with anything but limited recaptures, checks and promotion threats. But not everyone will agree. Some programs seem to thrive on extensions and seem to extend everything but the kitchen sink (at least according to their programmers!) If you base your testing on problem sets mainly you will probably conclude that many extensions are useful. But if you do self or auto testing you may (in my experience) find it less clear. Here are some extension ideas that I or others have tried ignoring singular extensions which is domain independent: Out of check, certain classes of them, extend if only 1 legal reply. Out of attack. Moving an attacked piece, interposing. Capture moves, (not necessarily recaptures) capture moves that restore the static eval to something close to alpha. Last queen capture! The idea being it's impossible to blow up the search and to get a better feel for the ensuing endgame. I don't really believe this is good enough reason to include it but who knows? Attacks. I prefer the out of attack extension because it is more directed. One program seems to pay special attentions to knight upattacks and forks. Combinations. Experimentation has been done on certain combinations of moves, like "extend the move following a check if it's a capture by the originally checking piece!" or "captures by pieces that just moved" etc. Most of these ideas are based on solving some problems noticed in games. The recapture algorithm for instance tries to solve the problem of horizon moves where BxN requires PxB and delays the real issues by 2 ply. Also most combinations are heavily capture oriented so they tend to look great with the right capture extensions. It may very well be that with null move prunning there is less of a case for extending moves. I am currently exploring more ideas for prunning but like extensions, this is tricky business! Each programmer seems to find his own bag of tricks in what works and what doesn't and very often it's not clear at all. Chess programming is a "black art" with a little science mixed in! - Don
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.