Author: Tord Romstad
Date: 06:27:54 12/29/03
Go up one level in this thread
On December 28, 2003 at 20:56:06, Tom Likens wrote: >I'll give it a shot. Essentially, what I've been experimenting with is >aggressively reducing and/or pruning various branches of the tree depending >on the current evaluation of the position and the previous level evaluation. >If things don't seem to be improving (or getting worse) then I either... > >1) reduce the depth the branch is searched or >2) eliminate it altogether. > >So far I've restricted this rather severely to positions where neither king >is in check, no other extensions have been triggered, the last move played was >not a null move, non-winning captures, non-hanging pieces etc. You may also find it effective to be conservative about pruning and reductions at nodes where there has been an extension in one of the last three or four plies of the path leading to the position. The disadvantage of doing this is, of course, that the path-dependency of the search decisions becomes even bigger. >One of the problems I've identified is that two identical positions may >not be pruned/reduced in exactly the same way, if their parent scores differ. >Storing the score for these nodes in the hash table is potential trouble, >since transpositions may have different scores. > >Uri's idea of only using these nodes for move ordering may do the trick but >I haven't really evaluated it yet. More than likely in the next day or two, >I'll add some scaffolding to my code to catch these type of nodes and gather >some statistics. Right now it's mostly guesswork. A technique I have found useful during development is the following: Instead of actually doing any forward pruning or reductions, set a local variable that says "I would have pruned this move if forward pruning was enabled". If it turns out that the move fails high, print the position and the move to a log file. By studying the log file, you can identify cases which your pruning techniques fail to understand, and use this information to improve the accuracy of your pruning. 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.