Author: Tony Werten
Date: 00:35:46 12/14/01
Go up one level in this thread
On December 13, 2001 at 20:11:27, David Rasmussen wrote: >On December 13, 2001 at 18:15:57, Robert Hyatt wrote: > >>On December 13, 2001 at 17:15:16, David Rasmussen wrote: >> >>>It has been "proved" that a two-level table with one table being depth >>>prioritized (I don't remember how many percent of the table that should be depth >>>prioritized, optimally). Crafty uses this approach, with 1/3 being DP, if I >>>remember correctly. I think it was in Brucker's thesis about hashing in game >>>trees. >>> >>>/David >> >> >>"proved" is too strong a word. multiple probes (2 or more) certainly reduce > >Ok, I meant ""proved"", then :) > >>the size of the tree. But they also have a fixed cost associated with them >>(they consume memory bandwidth that is already scarce). 2-level is a >>compromise between (a) using more bandwidth and going slower and (b) doing >>only one probe which will make the tree a bit larger. >> >>Smaller trees are good, so long as the total search time is also smaller. If >>you spend _too_ much time reducing the size of the tree, the cost of reducing >>the tree size may well exceed the savings caused by searching a smaller space. > >I know it's a design choice and a compromise either way. But Breuker tested a >lot of different setups, number of probes, number of tables, depth prioritized >etc. etc. And he """proved""" that in general, the scheme that you use, >actually, is the """""best""""". As with most (if not all) icca papers they don't prove a thing. The term "state of the art program" is very populair in those papers, but hardly applicable. It should mean: a program that represents the current knowledge but looking at their branchingfactors (specially when they switch all enhancements of to prove something) that can hardly be taken seriously. If you have a big bf than improvements are working uite often. The Journal does contain a lot of good ideas though, but you have to try them yourself to see if they work in your program. Tony > >/David
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.