Author: Ernst A. Heinz
Date: 05:04:04 06/11/98
Go up one level in this thread
On June 10, 1998 at 16:33:26, Don Dailey wrote: >Is my understanding correct that you are always trying a null move, >regardless of the static score? Yes, I suspect many/most people do it this way -- "DarkThought" incl. >For me it seems to work best to >have a static score very close (or above) beta. The algorithm as >I once saw it in the ICCA journal used a half pawn margin. Up to which depth to you use the static scores for deciding about null-move trials? >I found it slightly faster to simply do a test search with a zero >width alpha/beta window around beta. The idea being I only care >if we get a beta cutoff on the null move search. I'm not using >the result of one of these searches to tighten the Alpha side of >the bound either. With MTD this is a non issue but I found >it works find even with conventional alpha/beta. Again I think many/most people only do the zero-window test ... >Depth-3 selectivity does not work for me, although others report >great results. I always get nice speedups but find the program >weakens just enough to measure. Even when I'm very careful about >endgames and special case things. Yep, same for "DarkThought". >My biggest improvement is to not use null move near leaf nodes of >the main search. A "global" static attack test seems to be a very >nice improvement. It must be tuned but not too strict. It makes >the search much faster and it benefits from the speed. I can >easily measure the improvement in testing. It turns out that I >actually pick up a few things null move misses although visa-versa >is also true. I can only do a couple levels of these, then null >move is superior for its ability to see longer range threats. > >The basic idea of our static routine (much of it designed by Larry >Kaufman) is that we notice the highest valued piece attacked, and >then assume part of its value will be returned, so if a queen is >attacked, we assume only 2/3 of a queen is attacked. This is a >major speedup and the assumption is very good for all the pieces. >There is a value that seems to be on the threshold of right and >wrong, and it can be found by noticing a sudden drop in problem >solution depth when you get too agressive (like 1/2 attacked value.) I must admit that I do not quite understand how you actually prune. Could you please elaborate on this? =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.