Author: Robert Hyatt
Date: 16:30:19 08/20/00
Go up one level in this thread
On August 20, 2000 at 18:51:21, Dan Newman wrote: >On August 20, 2000 at 18:15:48, Robert Hyatt wrote: > >>On August 20, 2000 at 16:13:22, Frank Phillips wrote: >> >>>Can anyone sketch out the singular extension algorithm. I found some general >>>information on the net, but nothing that helps me understand how to implement it >>>in a PVS alpha/beta search. Descriptions tended to mention only that it is >>>invoked when there are only a few good moves in a variation. Since alpha/beta >>>does not yield the value of moves other than the principal variation, I am not >>>sure what this means in practice. >>> >>>Frank >> >> >>The simple case is on the PV search. When you search the first (and >>hopefully best) move at each ply, you search the remainder of the moves >>with alpha-w, beta-w, where w is some window offset (say 1/2 pawn). If >>all the other moves still fail low, then the 'best' move is better than >>the remaining moves by at least 1/2 ply. You re-search the 'best' move >>one ply deeper. >> > >What do you do if you get a fail-high on one of those remaining searches? >Do you reset the window to alpha-beta and re-search that one, or do you >just assume it failed high only because of the lowered window? Do you give >up on doing the singular extension at this point? It seems like if >another move (than the PV one) turned out to be much better, that it could >be "singular" itself... > >-Dan. That is correct. The thing you want to prove is that move S is at least W better than any other move. You hope the first move you search is best. But that obviously isn't the case every time. If you get a fail high, then you perform the singular test between that move and the previous best move. If neither fails low on the new window, then neither is 'singular' and you quit testing... > >>Ie some programs search deeper if there is only one legal way out of check. >>Suppose there are three legal ways out, but two of them drop all kinds of >>material. There is only one "reasonable" move and singular extension will >>follow it more deeply than the other two 'silly' moves...
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.