Author: Tord Romstad
Date: 10:55:46 03/11/04
Go up one level in this thread
On March 10, 2004 at 14:20:29, Sergei S. Markoff wrote: >I think that MTD(f) or MTD(f)-like schemer are very perspective in >future. But in SmarThink it seems not to work. Because of SmarThink >1) Hasn't any mechanism to keep PV excluding hash-table; 2) >SmarThink hashing scheme isn't quite good to use MTD(f)-drive. >In critical case it frequently produces a larger tree than >current method. Anyway if your engine works well with MTD(f) >you can try some of my ideas. For example in some cases you can >try after first move fail low to search all remaining moves instead >of trying to determine exact score of the first move. Than you can >discover that the first move is the only move that has eval >= >starting bound. Sometimes it's a good idea to stop failing low and >try to determine move singularity. At vice versa, you can try to stop >failing high to find other instead of failed. Be sure that not >all zero-window searches are equal. Larger bound causes a faster >searching (it's obvious, but...). Interesting ideas, I will think about them. >About null-move. Now only one hint. You have seen, I think, that >in Crafty, for example, zero move search result is using only >if it >= beta. Why not to use this result if it >alpha, but <beta? >Interesting question? In MTD(f) practically usual zero-move works >exactly this way. In MTD(f), of course >alpha but <beta is not even possible. >It's interesting: I think that in Fritz at least we can see some >non-well-known modifications of null-move search. It frequently >avoids zugzwang in positions that is absolutely ununderstandable >for standart scheme. I have no idea what Fritz does, but I have a primitive technique which at least helps in some simple cases: I don't allow null moves in positions where the side to move has very few safe pseudo-legal moves. >SmarThink before regular null-move search is trying to search >zero move to Depth=1 ply. Only if it causes result >= beta, than >SmarThink trying to search null-move to a full width. Anyway it's >not solving a most of null-move search problems. This is another technique I have experimented with, but the last time I tried it didn't work very well. Tord
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.