Author: Joachim Rang
Date: 12:10:45 12/24/04
Go up one level in this thread
Hi Uri, On December 24, 2004 at 09:31:57, Uri Blass wrote: >On December 24, 2004 at 07:04:59, Joachim Rang wrote: > >>Hi, >> >>in the next hour Leo will add the new Fruit 2.0 on his website. The new version >>comes with source as usual. Binaries for Linux come (perhaps) later. >> >>The biggest news: Fruit is no longer a blitzer! Fabien made a risky decision to >>activate history pruning with rather aggressive values, so you will notice Fruit >>prunes away a lot and reaches good depth rather soon. This makes him most >>probably weaker in lightning or blitz, but will (hopefully) pay off in longer >>time controls. If you aren't satisfied with the blitz-performance of Fruit try >>deactivating history pruning. > >I think that history based pruning when used in the right conditions is not >counter productive at blitz. That might be right, but the way Fabien implemented it, it_does_ hurt at least with 2+1 on my AthlonXP@1540MHz. More on how Fabien implemented it see below. > >You can try movei at blitz with history based pruning and without it. > >My impression was that the main advantage of it is in blitz based on the last >time that I tested movei with and without history based pruning in test suites. > >You can disable history based pruning in movei by changing >limit_history_depth from 101 to -1 > >You can reduce history based pruning in movei by increasing selectivity from 7 >to bigger number but there is some pruning that will always remain so pruning >does not converge to 0 by increasing selectivity. > >The problem is that I have condition: >badvalue>(goodvalue<<selectivity) > >selectivity=7 means that you prune the move if the number of fail low is more >than 128 times the number of fail high(there are more conditions and it is only >one of the conditions). > >The problem is that this condition can happen also when goodvalue=0 and badvalue >is small(there are more conditions so pruning when badvalue<=4 and goodvalue=0 >is not possible but pruning when badvalue=5 and goodvalue=0 may be possible and >how much you increase selectivity is not important. > >Note also that big selectivity may cause the value of goodvalue<<selectivity to >be wrong. > >I divide goodvalue and badvalue by 2 everytime when the remaining depth is at >least 10 and badvalue>1000 > >practically if I analyze the opening position for some minutes I find that >always goodvalue<(1<<15) and I guess that practically in games I have always >goodvalue<(1<<25) so it is no problem with the default parameters. > > > >> >>The threshold of history pruning is set to 60%. > >Can you explain what it means? I can only forward what Fabien answered me on this question: "To simplify, moves are given percentage scores in a separate history table. They start at 100% and go up/down depending on how often they fail high during the search. At a given node, moves that have a score below History Threshold are reduced. Move scores are always between 0 and 100. I think they are close to 100 most of the time due to the way I update scores (this might be a mistake but is the price to pay to have a significantly-different implementation from that of Tord and probably other engines)." My feeling is, that 60% is already a risky value, since it clearly misses some key moves in testsuites (but find others do to higher depth). regards Joachim > >Source code is free so it is possible to look at the source code but I prefer >first to read some explanation in words. > >I also see no reason that the parameter must be the same for different time >controls and the program may use the time control and the hardware to calculate >the parameter. > >Uri
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.