Author: Christophe Theron
Date: 01:11:02 10/14/03
Go up one level in this thread
On October 13, 2003 at 16:04:54, Russell Reagan wrote:
>On October 13, 2003 at 14:26:23, Anthony Cozzie wrote:
>
>>I think it is definitely possible to tune a program for longer time controls:
>>fewer extensions and a larger eval. OTOH, I don't think we're talking hundreds
>>of elo points. Maybe 50. (relative performace at short and long time controls)
>
>As the time control approaches infinity, the strength difference between two
>engines becomes zero, because both engines would solve the game and play
>perfectly (theoretically, of course). I think this is why people claim that
>their engine plays better at longer time controls. The weaknesses in the weaker
>engine simply masked.
>
>I believe that you can find out which engine is better by playing many short
>games as opposed to fewer longer games. As you get into very short games (1/0),
>the relative importance of various strengths and weaknesses becomes skewed. For
>instance, poor time management can be fatal in 1/0 games, while it will only
>have a slight affect on the results of games at longer time controls.
>
>Different time controls can help you find bugs or weaknesses of different
>magnitude. For instance, in a 1/0 game, if you have a big weakness, it will be
>exposed by an engine that is more solid all around. That weakness could be time
>management or king safety or whatever. The weaker engine will not see what is
>coming in the shallower depth allowed in a 1/0 game, while the stronger engine
>will pick up on that weakness sooner.
>
>This is simple to see if you play 1/0 games between two engines. I played
>Ruffian vs. Crafty in a 1/0 match, and in the majority of games, Ruffian's score
>would jump up, and Crafty's score would do the same within 2-4 moves usually.
>That would make it pretty easy to pinpoint exactly where a potential weakness
>existed in Crafty, or any engine that was tested in this manner.
>
>I think that you need to test at several time controls however. 1/0 games will
>show you weaknesses, but the matches also seem to be pretty one sided in favor
>of the superior engine. This could make testing whether or not a change is
>indeed an improvement difficult. For instance, you might fix your poor time
>management and king safety, but your engine might still get the pulp beat out of
>it because of some other weakness. 1/0 games pretty much expose any weakness
>that exists, from my experience. So to test one change, longer games (longer
>than 1/0) could prove more useful. Different time controls will show you
>different things. However, really long time controls show you less and less as
>the time controls get longer, so there is some point at which there is no point
>in playing longer games.
>
>But anyway, I agree with Christophe that being better at longer time controls
>isn't exactly a good thing. It only means that the longer time controls hide the
>weaknesses of the engine, which does nothing to help you improve your engine,
>and if anything it probably hinders your ability to do so.
>
>Also note that this is really only applicable in regard to computer vs. computer
>games. Computer vs. human at long time controls is pretty interesting IMO.
>Especially correspondance time controls.
Well said Russell. I agree.
I don't understand why other people do not get it either.
Christophe
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.