Author: Robert Hyatt
Date: 18:03:52 10/16/02
Go up one level in this thread
On October 16, 2002 at 03:47:01, Vladimir Medvedev wrote: >In his online interview Mr. CrazyBird told that DB did not use nullmoving >because it is not compatible with singular extensions. Could somebody explain >what are "singular" extensions (which difference with general meaning of >extensions technique? Are singular extensions used only for one move in the >node, while "common" - for all moves from the node? ) > >Why are they incompatible with nullmoving? Deep Blue 2 used singular extensions plus a new idea they had. The idea is that if you can prove that one move is clearly better than all the other moves, by a significant margin, then that move is called "singular". And it should be extended. The reason for this is handling threats properly. If only one move is real good, it might be something like a "spite check" that prevents you from capturing or doing something else to me that is good for you. If you don't extend such lines, you let the horizon effect sneak in artifact after artifact into your search. Null-move is exactly the opposite. You use a shallow search to see if your opponent has any serious threat, by just not moving. If you let him move twice in a row, and he doesn't kill you, then your position is very good. Good enough there is little point in spending more time searching it. The two tend to hurt each other. I had little trouble making a reasonable SE version of Cray Blitz (both fail-high singular and PV-singular) that performed well. I tried it on Crafty and failed miserably. I tried a "cheaper" approach used by Ferret and it also didn't work very well for me although it was "closer" to doing ok than _real_ singular extensions. Several have played with it since, and to date, I haven't seen any implementation in Crafty that is better overall than the non-SE version. Whether the "issues" can be resolved or not is another question, but the two do interact in bad ways. SE needs a real search to see if a move is significantly better than the rest. Null- move uses a shallow search to prune things. The interaction is not always favorable and SE misses extending where it should, or extends where it should not, because of the shallow null-move searches reporting back with confusing scores.
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.