Author: Vincent Diepeveen
Date: 13:35:43 06/28/04
Go up one level in this thread
On June 28, 2004 at 14:16:23, Joachim Rang wrote: >On June 28, 2004 at 13:10:06, Anthony Cozzie wrote: > >>On June 28, 2004 at 12:50:23, Peter Fendrich wrote: >> >>>I'm doing similar things with GA right now. The difference between GA and >>>Simulated Annealing is probably of not very important here. >> >>I've never heard of it actually. >> >>>My approach is somewhat different and I can elaborate more about it later on. >>>One problem that you are probably aware of is the low number of games. I belive >>>that small changes in an evaluation term will give minor differences in >>>performance and 10 games will maybe not capture that. For instance, play a match >>>vs some engine. Make a change that you know is lowering the strength and play >>>the same match again maybe x times and see what you get. >>>I would suggest shorter timecontrols in order to get more games. >>>/Peter >> >>I was hoping to get some help running the tests, as I don't have any commercial >>engines, or even windows. So all I can do is test against Yace, Zappa, pepito, >>etc. The hope is to get 10x10 or 20x10 :) I think shorter time controls are >>much more random. At blitz search is 99.9% of the game. >> >>anthony >> > >I disagree. I see no reason to believe that in short time controls eval is not >important and you can test that for yourself by playing Crafty with complete >eval vs. crafty with pst only on 2+1 - you will see the difference. > >I once read here in CCC or on the Programmers Stuff page from Ed Schröder that >he thinks one can test evaluation changes even in short time controls whereas >changes in the search like extensions or pruning behave different on different >depth so one needs to test on longer time controls too. > >My personal experience is that eval tuning on fast time controls is possible and > is valid also for longer time controls, but of course that is only my limited >personal opinion and it may be an exception ;-) Yes you are the exception. But when software improves you'll change your mind i guess. Let me just quote Stefan Meyer Kahlen: "I see clear differences between testing with short time controls versus long time controls, in blitz definitely different things work than for longer time controls". >regards Joachim
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.