Author: Fabien Letouzey
Date: 04:47:02 03/25/04
Go up one level in this thread
On March 24, 2004 at 20:18:23, Robert Hyatt wrote: >On March 24, 2004 at 11:17:48, Fabien Letouzey wrote: > >>On March 24, 2004 at 10:54:31, Robert Hyatt wrote: >> >>>On March 24, 2004 at 07:21:04, Fabien Letouzey wrote: >>> >>>>On March 23, 2004 at 17:17:01, Russell Reagan wrote: >>>> >>>>>On March 23, 2004 at 16:18:24, Robert Hyatt wrote: >>>>> >>>>>>Hashing can cause odd things. >>>>>> >>>>>>For example, you ponder for an abnormally long time and finish (say) a 16 ply >>>>>>search. As you searched position X at ply=1 (not depth = 1 but ply =1...) you >>>>>>get a "fail low" and store (say) score <= XXX, draft=15. >>>>>> >>>>>>Your opponent makes a different move and you start over. When you reach >>>>>>position X, you get a hash hit and you "fail low" because of it, bit when you >>>>>>re-search, you can't use that old fail low hash entry and you are not searching >>>>>>deeply enough to see the 16 ply problem with the move, so you get a screwy >>>>>>score. >>>>>> >>>>>>There is no solution to this... except drop hashing... >>>>> >>>>>To be more precise, you don't have to "drop hashing" completely to avoid this. >>>>>For example, you could still use the hash table only for move ordering and avoid >>>>>the search instability. Of course, it is less effective then. Pick your poison >>>>>:) >>>> >>>>I do exactly that at PV nodes in Fruit, for exactly that reason. >>>>Please stop having exactly the same ideas as I do ;) >>>> >>>>Fabien. >>> >>> >>>That doesn't solve the problem at all. A non-PV move can have a fail-high >>>stored in the table. You fail high on the move then fail low when you can't >>>resolve it... >> >>Sorry I was talking about depth/draft inconsistencies. >>I don't use aspiration at all. >> >>Fabien. > > >So was I. You have only solved the problem along the PV. It _still_ exists >along non-PV moves just as I explained... > >this has nothing to do with aspiration issues... Yes of course it does not fix all the problems, I should have stated it. However I think I gain some "stability" (and a complete PV as a side effect) at the cost of a 1-ply search (sometimes more) all along the PV. For some reason, I did not consider turning hashing off everywhere in the tree :) The tradeoff in my design is that null-window searches can do what they want (forward prune, be inconsistent, etc ...), and the pv-node search will try to accomodate with that. Fabien.
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.