Author: Joachim Rang
Date: 01:51:14 06/09/04
Go up one level in this thread
On June 09, 2004 at 03:08:20, Aivaras Juzvikas wrote: >could somebody recommend some techniques, how to test small improvements of your >chess engine? >the only way i know to test whether a certain change is an improvement or not is >to play hundreds of games vs other engines and then compare your new version >score to the old one. this consumes so much time and anyway its good for major >improvements, but small ones usually aren't visible. are there better ways to do >this? meybe playing a series of fixed opening lines (2 games for each opening) >could allow me to reduce the ammount of games needed to get a reliable result. >another question would be regarding time controls of these tests. if i'm playing >for example 2+2 games, is the result as reliable as the one i get playing 20+0 >(same ammount of games). > >the changes im speaking about are for example added extensions/reductions, or >added eval knowledge, not the speed improvement with the result of same nodes. For eval changes you might use very fast time controls, like 2+1 or 1+1 and you will quickly reach the necessary games. One might play 200 games from 100 fixed starting positions against the default settings and if it indicates an improvement is possible test that in a gauntlet against different opponents. For search changes it's more complex since it affects the branching factor and bullet games may not be accurately reflect the behavior with long time controls. Especially tactical extensions work very well in blitz time controls and loose their advantage with longer time controls. But if you can find reductions/pruning which decreases the search tree _and_ your results on tactical testsuites are the same I think it is a safe bet to assume that you have increased the strength of your engine. 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.