Author: Dan Andersson
Date: 14:01:09 05/09/01
Go up one level in this thread
On May 08, 2001 at 19:56:04, Dann Corbit wrote: >On May 08, 2001 at 19:35:28, Dan Andersson wrote: > >Well, having said that, will you say if there are theoretically sound forward >pruning techniques beside proof number search? Well as long as you have any kind of mate searcher or tactical prover where you can control the run time behaviour in detail e.g. number of nodes or run time. I think that will suffice. Then you will need a way to use it wisely. A static function taking remaining search depth and the position is ok but it will be hard to get any higher predictive power, say over 50% success rate. A dynamic function taking into account the efficacy of previous tries maybe by a killer or history heuristic will get a better prediction rate, or so it seems right now. >Define "all positions already available." Surely it can't mean what it sounds >like, so I assume you are talking about small subtrees -- probabaly already >pruned (soundly!) at that. >;-) I dont think I really need all positions available in any real way, it would be sufficient to construct the parts of the state space that are needed when they are needed. > >Speaking of sound/unsound pruning techniques... > >As everyone knows, NULL move is theoretically unsound. Suppose that I use NULL >move with R=2 and (starting with ply 1) my R=2 algorithm always completes two >more plies than the non-NULL move version. In a case like that, the algorithm >is theoretically unsound as a function of plies, but not as a function of time. >So can we come up with more than one definition of theoretically unsound? > AFAIK a normal implementation of nullmoves will not be able to correct zugzwang by going deeper. But if we allow two nullmoves after each other the search will be correct but at a higher search depth. >[snip] Regards Dan Andersson
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.