Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: About learning

Author: Marc van Hal

Date: 07:19:00 01/31/02

Go up one level in this thread


On January 30, 2002 at 16:18:57, Harald Faber wrote:

>On January 30, 2002 at 12:04:52, Jonas Cohonas wrote:
>
>>On January 30, 2002 at 10:05:30, Harald Faber wrote:
>>
>>>On January 30, 2002 at 05:56:51, Jonas Cohonas wrote:
>>>
>>>>On January 30, 2002 at 00:58:04, Harald Faber wrote:
>>>>
>>>>Gandalf always get's better results when its .lrn file is filling up, atleast
>>>>that has been my observations with earlier versions.
>>>>
>>>>Regards
>>>>Jonas
>>>
>>>I am not very interested in how strong an engine is after a few hundred games
>>>but in how strong an engine is out of the box.

Then you might be prety disapointed in Rebel too.
regards
Marc van Hal
>>
>>But if those are directly related how can you not find it interesting??
>>You were then one saying you were surprised that G5 beat GT2 and i only offered
>>my opinion as to how that could happen.
>
>No, not possible, because after each match I deleted the learn files. :-)
>
>>Let's say that in order to find the "true" strength of say gandalf, you would
>>have to play a few hundred or more games to reach its full potential, why do you
>>not find that important, Fritz, Junior, Tiger and all the other opponents
>>already have a healthy dose of booklearn i assume, so would it not only be fair
>>if Gandalf was tested with the same conditions?
>
>
>There is no sense in testing Gandalf 50 games vs. Fritz 7 and then play another
>50 games vs. another opponent. Reason: After 50 games Gandalf has some learn
>values the new opponent does not have. Only Fritz 7 and Gandalf do. So Gandalf
>would have advantage. And the more games you play, the bigger the advantage over
>the last opponent Gandalf plays.
>BTW that is exactly what SSDF does and why I perform my own tests. :)
>Sometimes the results fit, sometimes they don't...
>
>
>>Also why do you think the author even made the booklearn feature in his engine?
>
>
>For playing (humans) of course. :-)
>But learning should be used in a special way if you want to test and compare
>something like strength. SSDF somehow tests the learning too. I don't and I
>don't care. I don't want to see how strong one program can be after learning
>from 500 games. I want to see how strong the engine is NOW, out of the box.
>Because this is more that what normal customers have. Out of the box, no
>500-1000 learned tournament games. Right?
>
>
>>Regards
>>Jonas



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.