Author: Robert Hyatt
Date: 05:58:09 06/19/00
Go up one level in this thread
On June 19, 2000 at 02:17:48, Oliver Roese wrote: >On June 18, 2000 at 21:43:30, Robert Hyatt wrote: > >>On June 18, 2000 at 16:45:26, J. Wesley Cleveland wrote: >> >>>On June 18, 2000 at 16:05:01, Robert Hyatt wrote: >>> >>>>On June 18, 2000 at 15:50:43, Gian-Carlo Pascutto wrote: >>>> >>>>> >>>>>Hi all >>>>> >>>>>In the discussion of the 'Scalable Search Test' thread with >>>>>Ed Schroeder I mentioned that MTD(n,f) has the nice property >>>>>of making a fail-high pretty constant over time. I.e. the >>>>>search does not blow up as it does in a normal PVS searcher. >>>>> >>>>>Unfortunately it seems that this does not help when moving >>>>>up a ply...it even seems that the results of the MTD'ers >>>>>are quite terrible. >>>>> >>>>>The following though occured to me, if MTD allows you to take >>>>>small steps in the score plane, what about using fractional >>>>>ply increments to take smaller steps in the depth plane? >>>>> >>>>>Many of the best programs have now switched to fractional extensions. >>>>>Thus, fractional search depth must make sense. >>>>> >>>>>Iterative deepening is one of the most important improvements to AB >>>>>search. Thus, it makes sense too. >>>>> >>>>>Still, the programs use whole ply's in their iterative deepening >>>>>search. Why? It would make perfect sense to step in smaller increments >>>>>too. I feel this can even give improvements in tactical situations, >>>>>where the fractional extensions are triggered. >>>>> >>>>>I'm interested if someone has ever done or tested this before. Did it >>>>>work? What were the results? >>>>> >>>>>If you happen to have a program which uses fractional extensions, please >>>>>try it, and let us know how it works out. >>>>> >>>>>-- >>>>>GCP >>>> >>>> >>>>I tried this a good while back, but never really liked what I was getting. It >>>>is certainly worth trying... if you use fractional extensions. If you don't, >>>>it won't do a thing. >>> >>>I had an idea about this. If you kept track of how many extensions you did in >>>the search, if you had an unusually high number of extensions the iteration >>>before, you could search the next iteration to a lesser depth, e.g. >>>next_depth = last_depth + k*(nodes/(nodes+extensions)) >>>where k is equal to or somewhat greater than 1. >> >> >>Now your task is to test that. Sounds at least worth some testing. >> >>:) > >Could you shortly give an idea, why this approach could be beneficial?! >Lots of extensions indicate lots of stuff happening. >Searching is designed to overcome tactical barriers. Why then >_reduce_ the search depth? > >Thank you >Oliver Roese He wasn't saying _reduce_ the depth. He was saying, instead, "increase it more slowly". In many positions, one more ply exposes a set of tactics that can make the search blow up. IE at depth=12, no extensions get triggered to speak of, as those lines get pruned away by alpha/beta/ But at depth=13, suddenly you see deep enough to set a score that prevents the extended lines from getting scrapped quickly. the 12 ply search might take 60 seconds to finish. The 13 ply search might take 60 minutes. A 12.5 ply search would be somewhere in between.
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.