Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Duplicates test...

Author: Andrew Williams

Date: 15:03:26 03/08/04

Go up one level in this thread


On March 08, 2004 at 17:40:48, Peter Berger wrote:

>On March 08, 2004 at 14:31:02, Andrew Williams wrote:
>
>>On March 08, 2004 at 04:32:22, Andrew Williams wrote:
>>
>>>On March 08, 2004 at 02:50:56, Peter Berger wrote:
>>>
>>>>On March 07, 2004 at 16:46:28, Andrew Williams wrote:
>>>>
>>>>>improvedpreferred 108.5 - 61.5 standardpreferred
>>>>>
>>>>>The previous match score was 158 - 119 standardpreferred before I stopped it.
>>>>>
>>>>>I would NEVER have imagined a result like it, and it has been VERY easy to do
>>>>>this.
>>>>
>>>>
>>>>Doesn't this sound a little too good to be true actually ?
>>>>
>>>>It might be worth to have a closer look at the game statistics. I fear there
>>>>might be just one or two lines were the new preferred book outperforms the old
>>>>one badly, and they get repeated often in selfplay as you have no learner.
>>>>
>>>
>>>I checked the first match for duplicates and there were none. I'll have to test
>>>the second one when I get back. Remember, I have the huge book to fall back
>>>upon. I also have some time randomization just after the end of the book.
>>>
>>
>>I ended the match when I got back from work this evening. There were a total of
>>607 games and the final score was:
>>
>>newpreferred +341 -134 =232
>
>Do you think this can be real? Just think of the rating difference this would
>translate to . PM should be ready to face Shredder the day after tomorrow the
>latest if this trend continues ;).
>

I don't think it translates into anything w.r.t. other engines?! Or at least, it
won't translate into *much* with other engines. It's only really an improvement
in self-play matches, I think. But it's still funny to watch.

>>
>>I tested for duplicates and found none.
>
>I worded this badly. I was thinking about some general line not duplicates that
>might influence results heavily. I don't use SCID, but I know it can do
>statistics on ECO codes as Dieter posted some in the past. Sample idea: PM used
>to play E99 regularly with black, with white nearly always winning in selfplay
>and gave up on this line with black after you changed the preferred book.
>
>With such an effect taking place even the huge amount of games might not be a
>sign of a real improvement at all.
>

I take your point, and I'll take a look at the games in detail. But I don't
think there's much doubt that the book is better than before. Whether it
translates into a noticeable improvement against other opponents is a different
matter altogether, I fear.

>Of course I hope we'll get the 100+ points improved PM instead, simply with some
>general opening lines changed. Some of PM's opponents probably start to shiver a
>little anyway ..
>

Well, I think that 100+ points would be nice to have, but if I ever achieve that
it's going to be a LOT harder than what I'm doing at the moment!!

Andrew



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.