Author: Vincent Diepeveen
Date: 07:16:06 06/22/98
Go up one level in this thread
On June 21, 1998 at 21:50:20, Robert Hyatt wrote: >On June 21, 1998 at 20:46:16, Vincent Diepeveen wrote: > >> >>On June 21, 1998 at 15:59:33, Moritz Berger wrote: >> >>>On June 21, 1998 at 14:26:38, Ernst A. Heinz wrote: >>> >>>>On June 21, 1998 at 13:08:40, Vincent Diepeveen wrote: >>>>> >>>>>You can fill 300MB easily. >>>>> >>>>>Especially fritz. >>>>> >>>>>If you don't clean your hashtable (and fritz doesn't do this), then [...] >>>> >>>>Vincent, >>>> >>>>You are the first to claim that "Fritz" does *not* clear its hash tables >>>>after each move. Up to version 4, "Fritz" definitely did it ... >>>> >>>>Where do you get your knowledge from? >>>> >>>>According to my own experiences when playing "Fritz 5" I guess that it >>>>still clears its tables -- the first iterations just do not seem to fly >>>>by as they should if the tables remained intact. >>>> >>>>=Ernst= >>> >>>I observed that Fritz doesn't neccessarily find its own move from a previously >>>computed very deep PV (e.g. 15 ply or something) in the next move very soon. So >>>not even move ordering at depth-2 ply seems to be used. I certainly don't know >>>for sure, but would like to know Vincent's source of information. Maybe it's >>>gonna be news even for Frans Morsch ... >> >>Indeed weird, where fritz3 seems to suffer from this, fritz5 seems to repeat >>itself >>over and over again. Weird. So fritz3 not cleaning and fritz5 does. >> >>So it has become even dumber, now relying on piece square tables. >>That explains quite a number of things. >>. >>>Moritz > > > >also note that there are "levels" of clearing the hash. For example, I >*never* go through and zap everything. I do, on occasion, go through and >clear the scores only, if something changes seriously enough at the root >that I change the eval. For example, once development is complete, at >the root, I turn off EvaluateDevelopment() and I have to "scrub" scores >from the hash table as they would then be wrong. But I leave the other >stuff (move, threat flag, etc) alone... Move that code to your evaluation and you NEVER need to clear anymore, except for newgame. In diep btw i evaluate everything in evaluation. I also clear for analysis level btw, but that has other reason such as wanting a possibility to easily test how much nodes it needs. The weird thing is that Frans Morsch recently told me that he's just doing leaf evaluations, and not depending on piece square tables (he hated some posts some and probably me too did in this group and rgcc), now next time i speak to him he'll has to answer this. Fritz5 seems to clear always. Quite weird for a program where so many years has been worked at and where the programmer claims that it is depending on leaf evaluation. A third reason could be the case of course: that under windows his hashtables are allocated by fritz5 interface, which always clears them without Frans wanting to. Greetings, Vincent Vincent
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.