Author: Marc-Olivier Moisan-Plante
Date: 03:05:35 09/04/05
Go up one level in this thread
On September 03, 2005 at 16:28:07, David Dahlem wrote: >I'm currently using the same positions for my bench test as Crafty. And i'm >thinking about changing to different positions. What criteria should i use in >selecting good bench positions? All middlegame position, or a mixture of >opening, middlegame, endgame, quiet positions, non-quiet positions, etc.? > >Also any specific positions would be appreciated. :-) > >Regards >Dave Hi David, First I want to say I'm no expert in this (testing). Are you a member of the Winboard forum also? http://www.volker-pittlik.name/wbforum If you are, I would recommend you search the archives with <nunn and test>: your second result should be a thread named "Best method to test?". There is plenty of explanations here, although they might not directly answer your question about which positions should be in the test. If I can say, I think that the is no perfect positions: even a test of 20 positions (so 40 games as each program gets white then black in turn) could be "biased". For example if there are a lot of opposite-side castling positions, it could be that an attacking program gets the better of it in this test, but if most of the positions are "quiet" then maybe the more positional program will prevail. Among other things to considers there is the time control (apparently 10min/mate on a 2Ghz is equivalent to a 20min/mate on a 1Ghz - from what I read), and if there is book learning or some other sort of learning (some program learn even out of the book). Also the opening book is, of course, something we should think about. As posted lately on this forum a bad book (i.e. one side is almost winning after the opening) will tend to drive the match result toward 50%. Good luck, PS If you find relevant information about this just let me know.
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.