Author: Robert Hyatt
Date: 07:16:01 07/23/01
Go up one level in this thread
On July 23, 2001 at 08:33:49, Martin Giepmans wrote: >When I tested my Program SpiderGirl I found indeed a lot of "odball behaviour", >as you call it. With slightly different setups (a little bit more pruning or >a little bit less) I got huge differences in performance, ranging >from 0.1 sec (depth 18) to several minutes (depth 32, if I remember correctly). >It reminds me of the chaos-butterfly in London that causes a storm in Tokyo. >Is there a better way to test hashtable-implementation? Butterflies are Fine of >course, but ... I don't know of a simple, scientific way to test hashing. Fine 70 is a good one as you want to look for odd things. IE at something around depth 18 or beyond, up to depth 26, you should find winning a pawn. You should not see any drops in the score nor different moves, once you have initially found that Kb1 wins a pawn. You should see the score jump at least once or twice as you reach depth 36 or beyond, and you might see it actually reach +9 or so if you search long enough. You just want to be sure it doesn't go up, down, you change from one move to another over and over, etc. IE most programs will probably like Kb2 (more centralization) until they stumble onto the fact that Kb1 wins a pawn. Once they find that, they should stick with Kb1 forever as any other king move can not win that pawn, ever... If you see the above behavior, you are doing fine. If hashing is broken, you will notice the problem pretty quickly. IE on a single cpu, 700mhz, crafty finds Kb1 at depth 18, .05 seconds. Depth 20 takes .08 to complete, depth 25 takes .23 seconds, depth 30 takes 3.78 seconds, depth 35 takes 8.32, etc...
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.