Author: Tord Romstad
Date: 09:39:37 08/05/05
Go up one level in this thread
On August 05, 2005 at 12:10:52, Robert Hyatt wrote: >On August 05, 2005 at 05:58:52, Tord Romstad wrote: > >>The idea is based on the simple observation that in a program with reasonably >>good move ordering, if the first few moves in a non-PV node fail low, it is >>very likely that the search will fail low at this node. Late move reductions >>take advantage of this observation by searching moves late in the move list >>with reduced depth. In most implementations, if one of the reduced depth >>searches return a fail high score, the move is re-searched with full depth. >> > >OK... Bruce Moreland and I played with this quite a bit a few years ago, but >there was certainly a significant risk... Yes, there definitely is a risk. It is not too difficult to make the technique relatively safe from a purely tactical point of view, but I think it is likely to hurt positional play somewhat. How big this problem is depends on the engine, of course. For me the problem seems utterly insignificant, and late move reductions are a huge win. Without a doubt, the reason is that my evaluation function is worse than useless. An annoying problem is that late move reductions makes the search considerably less stable. Different hash table sizes of tiny changes in the move ordering can sometimes result in wildly different results when searching the same position to the same depth. This problem is particularly big if history counters are used when deciding which moves to reduce. I think it is better (but of course much more difficult) to base the decisions on evaluation combined with static and/or dynamic threat detection. 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.