Strelka 5 by Yuriy Osipov

[Steve B]

Hi Robert
Read the article about you in this months issue of Selective Search

nice photos
beautiful child you have there ..your a lucky man..everything else pales in comparison to that which i guess you realize already
if you want to avoid all of these negative posts just take my previous advice
establish your own forum while Houdini is all the rave and let the members defend it for you

PhotoGraphic Regards
Steve

[gerold]

Hi Robert
Read the article about you in this months issue of Selective Search

nice photos
beautiful child you have there ..your a lucky man..everything else pales in comparison to that which i guess you realize already
if you want to avoid all of these negative posts just take my previous advice
establish your own forum while Houdini is all the rave and let the members defend it for you

PhotoGraphic Regards
Steve

Plus one.

Best,
Gerold.

[Laskos]

Strong result from Strelka 5 with the new x64 compile

Code: Select all

    Program                            Score     %        Elo    +   -    Draws

  1 Houdini 1.5a x64               : 1114.0/2180  51.1   3204   12  12   34.0 %
  2 Strelka 5                      : 1066.0/2180  48.9   3196   12  12   34.0 %

Only 8 +/- 12 (95% confidence) Elo points weaker than Houdini 1.5a.

Kai

What is the time control?
I guess it is something like 6 seconds per game+0.1 second per move otherwise you could not have time to get so many games.

If you can make another match between them at time control that is 10 times longer then it may be interesting.

I used 4x time control, 4s + 0.4s, ~1 minute/game.

Code: Select all

    Program                            Score     %       Elo    +   -    Draws

  1 Houdini 1.5a x64               : 2122.0/4120  51.5   3205    8   8   36.5 %
  2 Strelka 5                      : 1998.0/4120  48.5   3195    8   8   36.5 %

Strelka 5x64 is weaker than Houdini 1.5ax64 by 10 +/ - 8 (95% confidence) Elo points. A new x64 compile of Strelka will probably be available soon.

Kai

[Osipov Jury]

...both Jury's offer and my answer are kind of a private joke between us.

Robert

This is not joke. I sent to your email sources of Strelka 5. Waiting for your response.

[bob]

Strong result from Strelka 5 with the new x64 compile

Code: Select all

    Program                            Score     %        Elo    +   -    Draws

  1 Houdini 1.5a x64               : 1114.0/2180  51.1   3204   12  12   34.0 %
  2 Strelka 5                      : 1066.0/2180  48.9   3196   12  12   34.0 %

Only 8 +/- 12 (95% confidence) Elo points weaker than Houdini 1.5a.

Kai

What is the time control?
I guess it is something like 6 seconds per game+0.1 second per move otherwise you could not have time to get so many games.

If you can make another match between them at time control that is 10 times longer then it may be interesting.

I used 4x time control, 4s + 0.4s, ~1 minute/game.

Code: Select all

    Program                            Score     %       Elo    +   -    Draws

  1 Houdini 1.5a x64               : 2122.0/4120  51.5   3205    8   8   36.5 %
  2 Strelka 5                      : 1998.0/4120  48.5   3195    8   8   36.5 %

Strelka 5x64 is weaker than Houdini 1.5ax64 by 10 +/ - 8 (95% confidence) Elo points. A new x64 compile of Strelka will probably be available soon.

Kai

I am not sure I quite buy that 10 +/- 8. It seems closer to 10 +/- 16, because each individual rating has a +/-8 error. When you subtract, both error bars should be included.

[Houdini]

This is not joke. I sent to your email sources of Strelka 5. Waiting for your response.

Nope, didn't get anything at houdini@cruxis.be .

[Michel]

I am not sure I quite buy that 10 +/- 8. It seems closer to 10 +/- 16, because each individual rating has a +/-8 error. When you subtract, both error bars should be included.

I don't think so. The scores are not independent (if one of them goes up, the other goes down). So +/- 8 looks correct.

Even if the score were independent the error bar should only be multiplied by 1.4.

EDIT: Actually I am not sure. I thought that the way Bayes Elo is set up
is such that the +/- 8 really refers to the difference. But I am not sure.

[Laskos]

Strong result from Strelka 5 with the new x64 compile

Code: Select all

    Program                            Score     %        Elo    +   -    Draws

  1 Houdini 1.5a x64               : 1114.0/2180  51.1   3204   12  12   34.0 %
  2 Strelka 5                      : 1066.0/2180  48.9   3196   12  12   34.0 %

Only 8 +/- 12 (95% confidence) Elo points weaker than Houdini 1.5a.

Kai

What is the time control?
I guess it is something like 6 seconds per game+0.1 second per move otherwise you could not have time to get so many games.

If you can make another match between them at time control that is 10 times longer then it may be interesting.

I used 4x time control, 4s + 0.4s, ~1 minute/game.

Code: Select all

    Program                            Score     %       Elo    +   -    Draws

  1 Houdini 1.5a x64               : 2122.0/4120  51.5   3205    8   8   36.5 %
  2 Strelka 5                      : 1998.0/4120  48.5   3195    8   8   36.5 %

Strelka 5x64 is weaker than Houdini 1.5ax64 by 10 +/ - 8 (95% confidence) Elo points. A new x64 compile of Strelka will probably be available soon.

Kai

I am not sure I quite buy that 10 +/- 8. It seems closer to 10 +/- 16, because each individual rating has a +/-8 error. When you subtract, both error bars should be included.

No Bob, that's simple. In an individual match the error is exactly 8 Elo points 2 standard deviations (95 % confidence). If one compares two individual results in a rating list comprising many independent, individual matches, giving X Elo points for both engines total error in the rating list, the error comparing these two engines is X*sqrt(2). In a rating list of many independent matches, one engine having an error X and another Y, the error when comparing them is sqrt(X^2 + Y^2).

Kai

[Laskos]

I am not sure I quite buy that 10 +/- 8. It seems closer to 10 +/- 16, because each individual rating has a +/-8 error. When you subtract, both error bars should be included.

I don't think so. The scores are not independent (if one of them goes up, the other goes down). So +/- 8 looks correct.

Even if the score were independent the error bar should only be multiplied by 1.4.

EDIT: Actually I am not sure. I thought that the way Bayes Elo is set up
is such that the +/- 8 really refers to the difference. But I am not sure.

It's ELOstat, and +/- 8 does really refer to difference in an individual match.

Kai

[Osipov Jury]

This is not joke. I sent to your email sources of Strelka 5. Waiting for your response.

Nope, didn't get anything at houdini@cruxis.be .

I tried twice to send. From jury_osipov@mail.ru and jury_osipov@yahoo.com