Author: Robert Hyatt
Date: 06:45:20 11/24/97
Go up one level in this thread
On November 24, 1997 at 01:35:38, Komputer Korner wrote: >Now that we have your word that everyone uses some sort of replacement >strategy when the hash table fills up, doesn't this mean that Chris >Whittington's description of the program acting like a drunken sailor >when the hash table is filled, a bit of an inaccuracy, thus the results >of the Korrespondence Kup are valid? >On November 21, 1997 at 12:34:26, Robert Hyatt wrote: > >>On November 21, 1997 at 11:06:15, mike cooter wrote: >> >>>In Fritz 5 when I set it up to think on infinite level, once my >>>hashtables fills up to100 percent, is the computer still thinking on >>>what is the best move? If it is, then what is the hash tables for? >> >>they are used to speed up the search. However everyone uses some sort >>of replacement strategy when the table fills up, so that it continues to >>help the search, although at a somewhat reduced effectiveness... As I've said all along... long searches certainly run into a sort of "wall" where the branching factor seems to go up by some amount... But it doesn't mean the programs play blindly or stupidly... I don't know who uses what replacement strategy. Don Beal wrote a paper a few months back testing the algorithm I used in Cray Blitz vs a simple two- level replacement strategy like I now use in crafty. For cases where saturation became an issue, the old Cray Blitz algorithm was better. But we aren't talking about 10X faster, or 5X faster, which is a common misconception. Someone posted some results here that showed Crafty slowing by 2x when it "hit the wall". 2x is significant, but hardly critical at correspondence depths...
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.