Author: Tord Romstad
Date: 08:07:02 02/13/04
Go up one level in this thread
On February 13, 2004 at 09:23:38, Bob Durrett wrote: >On February 13, 2004 at 04:18:34, Uri Blass wrote: > >>On February 13, 2004 at 01:47:40, Tord Romstad wrote: >> >>>I'm beginning to get a nagging feeling that recursive null move pruning (at >>>least the conventional kind) is possibly the *worst* thing to happen to computer >>>chess ever. There has to be a better way to reduce the size of the tree, but >>>I have no clue how to find it. :-( >>> >>>The zugzwang problem is not too serious, of course, and if you really care >>>about it it is not hard to solve. The real problem with recursive null move >>>pruning is that it performs horribly at finding long non-forcing lines. For >>>instance, a human player could take a quick look at a position and see that >>>black needs to exchange off white's strong knight on c4, and notice that this >>>could be achieved by the maneuvre f7-f6 followed by Bh7-g6-e8-d7-c8-a6xc4. >>>A recursive null move searcher needs a huge search depth to find such plans. >>> >>>Tactically, recursive null move pruning performs really well. Strategically, >>>it's horrendous. >>> >>>Tord >> >>I believe that chess is mainly about tactics and you can decide about rules when >>not to use null move pruning > >That sounds like the germ of a super new chess algorithm! The essence of your >idea seems to be that various tree shaping strategies/methods would be selected >based on intelligent evaluations of positions. For certain positions you might >do something, for tree shaping, which is entirely different from what might be >done for another position. It is not new at all. We all do things like that. >>(if it seems that the moves are a plan). >> >>I do not do it but it is one of the things that I should try and the problem is >>to have a function to decide if some moves are a plan. >> >>In most practical cases at least part of the moves of the plan threat something >>so you do not need a huge depth. Of course it is possible to patch some of the holes by adding exceptions and code for many different special cases, but to me it doesn't seem like the right thing to do. I also find it hard to believe that an algorithm which uses so little domain-dependent knowledge (the only domain-dependent knowledge used in recursive null-move pruning is that zugzwang is rare in chess) is anywhere near optimal. That several of the top commercial programs do not use null move (at least not in the conventional way) is further evidence that it is possible to come up with something better. Tord
This page took 0.04 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.