Author: Peter McKenzie
Date: 19:34:53 03/03/01
Go up one level in this thread
On March 02, 2001 at 23:07:49, Ernst A. Heinz wrote: >>>>Why does it make sense? I have not yet tried it but I don't understand why >>>>should it be better? Do you have any opinions? I allways want to understand the >>>>features I put into my engine and I don't get this one... >>> >>>Easy -- the relative cost of a depth-reduced null-move search >>>at frontier nodes (with a remaining depth of 1 ply) is just >>>too high. Most normal searches from there also go straight into >>>quiescence mode and the gain of null-move cutoffs does not seem >>>to outweigh the added null-move search effort. >>> >>>=Ernst= >> >>I tried this for a quick test. It is actually slower overall than doing >>null-move everywhere. A couple of positions were faster. But more were >>slower. I ran all the kopec positions... > >As usual, your mileage may vary and it does not work for >all programs. Do you apply futility pruning at frontier nodes? > >If not, this may explain your negative results because you are >much less selective at frontier nodes then which increases the >chances of the null-move overhead to pay off. This sounds like a plausable explanation, but I suspect that the application of futility pruning (FP) at pre-frontier nodes could be a significant factor here too. For example, if it is a fail low pre-frontier node and I don't apply futility pruning then there will be a number of frontier node positions searched that will likely easily be pruned by nullmove. These positions would have been avoided altogether if FP were used at pre-frontier nodes. So my theory is that the application of FP at pre-frontier nodes would decrease the percentage of nodes that are successfully pruned by nullmove at frontier nodes. Thus, nullmove at frontier nodes becomes less worthwhile when you do FP at pre-frontier nodes. Whether the usefulness of nullmove at frontier nodes decreases enough to make it not worthwhile, I can't say. Although this thread would suggest that is the case. Of course I haven't tested this stuff for ages... Must look at what my program does! Peter > >=Ernst=
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.