Author: Roland Pfister
Date: 23:51:00 11/20/97
Go up one level in this thread
On November 20, 1997 at 14:35:11, David Eppstein wrote: >I thought I would share with the group a technique I tried yesterday, to >get the program I'm working on to play better in a class of positions >where it was having horizon-effect problems. > >I identified a class of "suspicious moves", moves that are a priori >unlikely to be correct but that are reasonably likely to be used as >delaying tactics in the horizon effect (and that were in fact occurring >in the horizon effect problem positions I was looking at). The moves I >classified as suspicious were non-captures when a seemingly good capture >is available. The move ordering I use naturally places these moves later >than the captures, as I think would be true in many or most programs. > >Anyway, if I search a suspicious move and it appears worse than >previously searched moves at the same node (i.e. below alpha), I trust >that result. But, if it comes in greater than alpha, I immediately >become suspicious and search the same move to one greater ply. The >deeper search result then becomes the value of that move. > >The intention is that this extension doesn't blow up the search tree >size very much (because suspicious moves are rarely chosen as best) but >improves the search accuracy. It is hard to be objective, but my program >does seem to be playing better after this change. > >Is this technique common in other programs? Can anyone suggest >refinements that would make it work better? I tried the following extension: if the hash move is a move that does not capture the last moved piece even if it would be possible and would be good (statically) and it doesnt capture a better or equal piece, extend one ply. That was intended to work in positions where one side puts a piece en pris and has a threat so that the opponent can't capture that piece. I discarded it because it made my tree bigger and the search slower( in Bratko-Kopec ) without getting better results. I will try your idea. Let's see what happens. Roland
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.