Author: José Carlos
Date: 01:10:01 08/21/00
Go up one level in this thread
On August 20, 2000 at 20:12:02, Uri Blass wrote: >On August 20, 2000 at 18:44:04, Georg v. Zimmermann wrote: > >>On August 20, 2000 at 18:29:06, José Carlos 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. >>>> >>>>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 suggests me an idea, I don't know if it has been tried: >>> >>> Suppose you've searched all moves at depth d. Then, at depth d+1, the PV move >>>has a bigger value then at d. In that case, you could simply go to d+2 without >>>looking at the rest of the moves. >>> Only when the PV value drops down at any depth, search the rest of the moves. >>> >>> Just a thought... >>> >>> José C. >> >>This seems flawed to me >> >>1.) Lets say at ply7 you think position is equal at ply8 the first move searched >>turns out to win the queen, but the second would mate ? Why skip it ? > >If both moves win then it is not important. >The important case is when you have not a win with move A and have a win with >move B. > >I do not say that the idea is practically good but I understand the reason for >the idea. > >If the pv has a bigger value the chances to find a better move are smaller so >instead of wasting time in searching the rest of the moves it may be better to >save time by not searching them. > >The problem with the idea is that it is also correct that when the pv has bigger >value you search less time the rest of the moves so the gain is also smaller >from not searching the rest of the moves and not only the loss(in cases that >another move is better. Yes, exactly. But what I was thinking about is that, if the PV value increases (assuming you have a 'good enough position'), the rest of the moves are not likely to give a big improvement, but if one of them is, say 0.03 better, then you fail high and have to research with a wider window, just to find you have a very small improvement. Probably this idea only makes sense in won positions, or positions with a clear advantage, where you want to make sure the move you are gonna play doesn't have a hidden trick. Then, you go much deeper with that move than with the others. If nothing strange happens, you can play it and you know you're gonna win. There can be a better move, of course, but you'll win anyway. José C. PS.: I'll try it in Averno when I have some free time :) >>2.) If I go to d+2 with move1 immediately the other moves will not get better >>sorted. >> >>3.) The first move takes longer to search than the others anyway. > >It is not always the case but usually it is the case when the pv gets a bigger >value and it is the problem with the idea. > >There are some cases when programs search the first move for a short time and >waste almost all the time in other moves(mainly when there is a transition to a >simple endgame) and I think that in this case it may be a good idea to search >the first move to a bigger depth than the other moves. > >Uri
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.