Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Reproducibility of Nunn matches is not so good

Author: Bertil Eklund

Date: 13:53:15 05/02/00

Go up one level in this thread


On May 02, 2000 at 08:14:58, Ulrich Tuerke wrote:

>On May 02, 2000 at 07:39:03, Bertil Eklund wrote:
>
>>On May 02, 2000 at 03:48:48, Ulrich Tuerke wrote:
>>
>>>On May 02, 2000 at 03:41:22, Christophe Theron wrote:
>>>
>>>>On May 02, 2000 at 03:22:36, Ulrich Tuerke wrote:
>>>>
>>>>>On May 02, 2000 at 01:10:58, Jouni Uski wrote:
>>>>>
>>>>>>About one week ago I posted "sensational" blitz Nunn test result Crafty17.10 -
>>>>>>Fritz6a 12-8. And I wasn't lying! Now I have more time to re-check and second
>>>>>>match ended 9 - 11 for Fritz. I also repeated two other matches with interesting
>>>>>>results:
>>>>>>                           1.              2.
>>>>>>Crafty 17.10 - Lg2000   13 - 7         9,5 - 10,5
>>>>>>Fritz6a - Lg2000        13,5 - 6,5     11 - 9
>>>>>>
>>>>>>Exact enviroment: AMD 450Mhz, ponder of, 16+16MB hash, 4m+1s level, 3+4+some
>>>>>>5 piece TBs, Fritz6 interface, early resign.
>>>>>>
>>>>>>Conclusion: after 20 games You don't know much yet...
>>>>>>
>>>>>>Jouni
>>>>>
>>>>>With learn off, the games should be exactly reproducable (IMHO). Why should the
>>>>>search algo of either prog under the same pre-conditions produce another best
>>>>>move for any of the positions some time later?
>>>>>If you are right, then IMO either (or both) progs are kind of buggy, accessing
>>>>>some non-initialized data or similar ?
>>>>>Or does any body have anothe explanation ?
>>>>>
>>>>>Uli
>>>>
>>>>
>>>>The main problem is the way time is measured.
>>>>
>>>>On the PC, the time functions only return multiples of 1/18.2 seconds. Even
>>>>functions that are supposed to return the current time in 1/1000 of s are not
>>>>accurate to the millisecond. They are accurate to approximately 5 hundreds of s
>>>>(1/18.2=0.054945...).
>>>>
>>>>So depending exactly when you started a search inside a 1/18.2 seconds time
>>>>slice, searching exactly the same number of nodes could fall randomly just
>>>>before or just after another given 1/18.2 time slice.
>>>>
>>>>So when you measure the time taken by a given search, always the same, you end
>>>>up with pseudo random results. A search that takes exactly 1 second can be
>>>>measured at 0.989s, or at 1.044s, and it depends if it started just before of
>>>>just after a clock tick.
>>>>
>>>>The time allocation algorithm of a program could decide that if a search takes
>>>>less than 1 second, it will allow it to complete the next iteration. If it takes
>>>>more than one second, it will stop the search immediately and play.
>>>>
>>>>In the case of our 1 second search, it will sometimes stop the search and play
>>>>after 1 second, and sometimes let the search continue for longer.
>>>>
>>>>There is no way to avoid this problem. Even with a millisecond-accurate timer.
>>>>Even if you can measure the time up to the microsecond.
>>>>
>>>>Random events such as mouse moves, hard disk saving mode and autoplayer random
>>>>lags only make this problem worse.
>>>>
>>>>This is not a bug. You can call this a "quantic" problem. :)
>>>>
>>>>
>>>>
>>>>    Christophe
>>>
>>>Thanks, sounds reasonable to me. So, in particular at very short time controls,
>>>I would expect that these effects could have some influence in blitz games where
>>>the uncertainty is relatively large compared to the time which is needed to
>>>search a root move, and Jouni played at rather short controls. I do not think,
>>>that these "quantum uncertainties" would be that drastic at tournament controls.
>>>
>>>Well, if Heisenberg knew ?
>>>
>>>Uli
>>Hi!
>>
>>I have played a few 1000 games with Comet (most in DOS). There is two programs
>>that never seems to repeat a lost or won game and it is Comet and Nimzo.
>>I thought it had something to do with the hash-tables. Or do you you use some
>>random element to choose moves? Rebel always plays the same if the conditions
>>are equal but never Comet or Nimzo.
>>
>>Bertil

Hi!

They never repeat wins either.

Bertil
>
>That's not a surprise, because Comet has a positional learning implemented,
>using a file on disk.
>I think, that I have heard that the same is true for Nimzo.
>These are file with the extension .lrn if I remember right.
>
>Uli



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.