Author: Gerd Isenberg
Date: 13:01:43 12/09/03
Go up one level in this thread
On December 09, 2003 at 06:54:24, Ulrich Tuerke wrote: >On December 08, 2003 at 16:19:48, Tim Foden wrote: > >>On December 08, 2003 at 11:02:22, Alvaro Jose Povoa Cardoso wrote: >> >>> >>>I've been thinking about the efficiency of the history heuristic at high search >>>depths. >>>It seems to me that the history table will be overwritten many times if we have >>>a search of several billions of nodes. Additionally, as the search moves to >>>different parts of the tree the history table values will be somewhat trashed. >>>What do you think we could do about this? >>>Maybe limit the history heuristic to a certain depth (ex: the nominal depth). >>> >>>Comments anyone? >> >>What happens in GLC is whenever it increments a value in the history table it >>checks it against a maximum. If the maximum value is exceeded, it divides all >>values in the table by 2. > >Isn't this a bit expensive ? I guess that this may happen several times per >iteration, and perhaps very often at higher iterations. > >I do it like this > > if (*hisptr < maximum) > *hisptr += ((depth*32) >> iteration_no); > >"depth is the distance of the current ply to horizon, so this measures in some >way the size of the sub-tree which had been cut. > >i.e. the increment is smaller at higher iterations, because at higher iterations >I expect the history to be filled much faster. >At start of a new iteration, I divide all entries by 2. > >However, I'm not sure how well this works. > >Uli >> >>Cheers, Tim. Hi Tim and Uli, i do it in a similar way than Uli but increment some common overflow counter for each color in an else case. If the number of overflows per side exceeds some threshold i divide all history counters (unsigned chars) of this side by two (or four) and clear the overflow counter. Trade off of saturation versus too many divides and a few more tuning screws ;-) Cheers, Gerd
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.