SWCR: SF 2.1.1 JA x64 PHQ & SF 2.1.1 JA x64 ...

[Uri Blass]

If we assume 70 elo for doubling the speed then
10% speed difference can give more than 7 elo
1.1^7.3=2.005241
70/7.3~9.589 so by simple math 70 elo for doubling gives more than 9.5 elo for 10% speed improvement.

of course 2.3% is a different story and
I can expect it to give only 2.3 elo rating improvement.

The question is if we really have only 2.3% speed improvement for the IPON.

I do not know if it is possible the speed improvement is bigger when you give stockfish more time and the time for the bench result of stockfish was not enough to show a bigger improvement than 2.3%

[IWB]

Assuming 70 with double speed (I tend to 60) you are right with 9.5 Elo, the damed logarithm comes in

Anyhow, going from 2.1 to 2.1.1 I had a "jump" of 10 Elo. As I have 2.3% speed difference on the same hardware the much more logical thing I see is error bar than the doubtfull speed increase.
More games will tell ...

Bye
Ingo

[Frank Quisinsky]

Hi Uri,

IPON have different other results if you compare it with SWCR. A good example is Junior 12.5.0.3 ... around 30 ELO, Rybka 4.1 and much others.

I think this have to do with the time control. A bit more as the double time I used.

Important are that I am playing without resign mode. Many interesting statistics with the available SWCR database are possible.

In the maintime I played more games with SF 2.1.1 PHQ-2 ... it could be possible that we get +25 ELO to default. I will test it official in SWCR with ponder = on and the best PHQ result I will hold for my SWCR. The result from the weaker setting I will delete.

I think I start this test with Stockfish 2.1.1 JA x64 PHQ-2 and Hannibal 1.1 x64 to the same time after Komodo 2.03 JA x64 / Equinox 0.97e x64.

Interesting is what the SF programmer wrote. I don't know that SF is more optimated for Intel processors, completly new for me. But this explain 2-3 ELO.

And a explanation for the other ELOs are quiet clear for me.

Positional programs are stronger in blitz, found good and important positional moves faster. Tactical strong programs need more time for positional moves. With more time the tactical programs are stronger. This explain the differents from IPON to SWCR very easy.

Most are looking on:
How many games an engine plays! But more important is ... how many participants are used !!

50% - 50%
Both are important for a good rating!

1. Many different opponents
2. Many games too

The different from 1.000 - 2.000 games = 0.3 ELO points more exactly.
That's all ... could you test in self work with the available SWCR database.

Have a nice day!

Best
Frank

[mcostalba]

In the maintime I played more games with SF 2.1.1 PHQ-2 ... it could be possible that we get +25 ELO to default.

I have done a test with PHQ-2 (mobility at 115-85 and aggressivness at 135-85) and on self test it is more or less same ELO than standard. I have also tested only aggressivness-cowardice change and it seems a bit weaker than standard and also I have tested only mobility changes alone and it seems equal or even a bit stronger. So at the end I have decided to committ the full PHQ-2 settings becuase I want to trust your very important test that is done at longer TC and with many opponents.

Interesting is what the SF programmer wrote. I don't know that SF is more optimated for Intel processors, completly new for me. But this explain 2-3 ELO.

SF is not optimized for Intel. What happened is that Jim has an AMD CPU so when he tested Microsoft MSVC vs Intel compilers on AMD it seemed that MSVC was a bit faster, then, eventually we found thanks to an user report, that on Intel hardware Intel compiler was faster and so Jim compiled again SF 2.1.1 but with an Intel compiler even if the result was a bit slower on his AMD, but I can verify it is faster on Intel i5 CPU, the one that I have.

[Frank Quisinsky]

Hi Marco,

OK, test will be start in SWCR (PHQ-2) Tuesday, perhaps Monday in the evening ... I have to wait of the final results from the still running test of Equinox and Komodo.

Shortly:
I test also (500 games without ponder only) PHQ-2 with 40 in 10 and 40 in 5. The result is ...

40 in 10 made 1.4% more vs. the same participants = around 15 ELO as 40 in 5. It seems that the PHQ settings are good for longer time controls. 500 games are not enough, I know.

Interesting what you wrote ...
Mobility middlegame is my idea in PHQ. This one I think could be good after my database analyzes with the latest 4 Stockfish versions.

The question is now after your and my test ...
What I should test exactly?

1. PHQ-2
2. An other setting you like ... perhaps with mobility settings only?

Both to the same time?
Give me a second hint, we can create a SWCR-Stockfish-Test-Weak
Why not!

Thanks for your explanation to the compiles.
Now quiet clear for me.

Best
Frank

[mcostalba]

The question is now after your and my test ...
What I should test exactly?

1. PHQ-2
2. An other setting you like ... perhaps with mobility settings only?

PHQ-2


I have just done self-test and anyhow even with mobility only results are much different than what you found in your tournament, so I think your test can really shed new light and add new elements above the usual self test. So I think it is better to test it with full settings mobility+king tweaks.

[Frank Quisinsky]

Hi Marco,

OK, PHQ-2 only.
I will test it vs. the same participants, I used for PHQ-1. My new tester from Italy test PHQ-1 too vs. two older IvanHoe versions. Here the results are not very good. First result I added in database and ELO calculation, second I got today. I am not at home and can update the rating list monday.

After my test without ponder so far its 19 ELO more as default for PHQ-2. Let us see with SWCR conditions now

Best
Frank

[Frank Quisinsky]

Hi Uri,

after my calculation from all TOP-30 engines = double speed = 54 ELO.

Best
Frank