Author: Gerd Isenberg
Date: 10:33:50 05/25/04
Go up one level in this thread
On May 25, 2004 at 07:59:33, Tord Romstad wrote: >In all implementations of iterative deepening I have seen, the search depth >is always incremented by exactly one ply between each iteration. Is there any >reason to believe that this is optimal? Has anybody tried other increments? > >I am now running some tests with an increment of 3/4 of a ply. So far, the >results don't seem very different from the usual 1 ply increment. If it turns >out that increments different from 1 ply are no worse, this could perhaps be >useful in time management. The engine could use 1 ply increments most of the >time, but occasionally add only half a ply if just a small amount of the >allocated thinking time is left. > >Any thoughts about this? > >Tord Hi Tord, yes it is worth to play a bit around with it imho. What about feeding back some statistical information from previous iterations, e.g. branching factor to vary the increment? A kind of closed loop to keep the branching factor constant, considering possible odd/even effects ;-) What about to count at the leaves how often extensions results in additional plies and even to count how often an additional ply is missed by some margin. Those counters may advise how likely an "explosion" and bigger bf in the next iteration is, and whether the increment should be better reduced. Timemanagement (including ponder hit handling) is a big issue and my one is far from being perfect. Beside remaining time and moves there are a lot of other time managemant heuristics. One should consider "easy moves" like obvious recaptures with no change of the best move so far over all iterations, exceeding some relative duration threshold of that best move so far. May be some conditional singularity test of the root moves at some fixed depth. The score and the graph of scores over previous itereations of this position is interesting as well as the "fliprate". Determing some heuristic root position properties, eg. wild/tactical versus quite may be considered. Beside that, analyzing root moves a bit more to get an idea of dubious/interesting moves and sacs and to take some more time if best move changed during last iteration from none dubious move to a dubious one. Same may be true after analyzing the PV. Much room for further research and improvements. 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.