Author: Vasik Rajlich
Date: 01:40:07 04/02/04
Go up one level in this thread
On April 01, 2004 at 10:03:36, Renze Steenhuisen wrote: >On April 01, 2004 at 09:43:34, Renze Steenhuisen wrote: > >>>The main thing which will drive up your first cut rate is internal iterative >>>deepening. (If you have it implemented and are getting 60-80%, then you have a >>>bug.) >> >>Can anyone give me typical numbers on this point? >> >>Thanks! > >Putting Internal Iterative Deepening in as follows, doesn't give me any better >FH-% numbers. What does happen, strangely enough, is that finding a Fail-High on >a move is postponed one ply. THIS SEEMS WEIRD TO ME? > >/**** IID ****/ >if( tt_move==NO_MOVE && remain_depth>PRE_PRE_FRONTIER && try_null_move==TRUE ) >{ > value = pvs_search(tree, alpha, beta, remain_depth-2, DO_NULLMOVE); > if( value<=alpha) > value = pvs_search(tree, -MATE, alpha+1, remain_depth-2, DO_NULLMOVE); > lookup_tt_entry(tree, depth, &tt_value, &tt_move); >} > >Cheers! > >Renze I don't understand what you mean "postponed by one ply". Re. #s for IID, I don't have them for failing high on first move. Actually, failing high on first move is not the full story. You could fail high on first move, but with an inferior move, leading to an bad fail-soft return value. Later, you will need to re-search. (This is for MTD (f).) Move ordering changes can be conveniently and accurately tested by looking at nodes needed for a certain depth (in the case of IID, which spends nodes in order to reduce node counts) or time needed for a certain depth (in the case of static ordering). IID reduced my trees very clearly (something like 7 or 8%), and depth reduction 1 was best by a tiny margin. Note that differences in scores when the move ordering changes can be ignored. The exception is when you have very coarse eval granularity. Vas
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.