Author: Peter McKenzie
Date: 10:31:25 01/12/01
Go up one level in this thread
On January 12, 2001 at 05:33:16, David Rasmussen wrote: <snip> >>I guess the basis for some successful pruning methods is figuring out that a >>position is really great, and that it is a waste of time searching it to the >>normal depth. This is all that null move pruning does after all. >> >>So how can we improve on null move pruning? >>Well, there are two areas that come to mind: >> >>1) find a cheaper way to determine if the position is great. The null move >>search is quite cheap, but it certainly isn't free. >> >>2) find a more accurate way to determine if a position is great. I suspect that >>the null move search is still quite pessimistic and doesn't prune alot of >>positions that it should (of course the flip side is that it doesn't make too >>many mistakes either!). >> >>The cheapest way to estimate if a position should fail high is to just look at >>the static eval. Not very accurate though! Perhaps you can use this to guide >>your search? Thats what fail high reductions do I think. If you could combine >>that technique with null move pruning, then you might be on to something. >> >>Maybe you can make better use of the eval function by looking at its component >>parts? For example: >> if eval > beta + margin && ourKingSafetyPenalty < xxx then >> prune(); // or reduce search depth or whatever >> >>Anyway, just a few random ideas :-) >> > >Nice ideas. > >But I was looking not so much for new ideas, as for the already known forward >pruning techniques that a program such as GT might be using or at least build >on, to achieve such "good" performance, node-wise. The problem is that there is very little published on forward pruning apart from futility/razoring, nullmove and FHR! It seems to be mainly the commercial authors who use other techniques (or improved versions of the standard techniques), and they don't usually give away any secrets... Don Dailey did write up a few ideas in this forum quite some time ago if you want to search the archives.
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.