cm11: stunning result for selectivity =21

[ozziejoe]

sel 21 gets a +60 pnt boost in this particular test over the default cm. I have never seen a default parameter change make this big of a difference. Even if we assume that cm_sel21 got a bit lucky, it still is likely to be notably better than the default settings


Doesn't the author of cm do any testing?

all necessary info below


5/01/2008 1:55:16 PM :

Program Elo + - Games Score Av.Op. Draws

1 Rybkav2.3.2a.w32 : 2584 52 51 150 77.7 % 2367 28.7 %
2 Fruit-051103 : 2457 50 49 150 62.7 % 2367 25.3 %
3 kinguci_sel21 : 2433 34 34 300 50.5 % 2430 26.3 %
4 EngineShredder10UCI : 2431 46 46 150 59.0 % 2367 32.7 %
5 Spike1.2 : 2431 46 46 150 59.0 % 2367 34.0 %
6 Naum : 2379 45 45 150 51.7 % 2367 34.0 %
7 KingUCI_dual : 2366 33 33 300 40.8 % 2430 32.3 %
8 kinguci_single : 2307 34 34 300 33.0 % 2430 31.3 %
9 Ruffian_210 : 2307 49 49 150 41.3 % 2367 25.3 %



3 kinguci_sel21 : 2433 300 (+112,= 79,-109), 50.5 %

Rybkav2.3.2a.w32 : 50 (+ 9,= 14,- 27), 32.0 %
Fruit-051103 : 50 (+ 17,= 10,- 23), 44.0 %
EngineShredder10UCI : 50 (+ 16,= 17,- 17), 49.0 %
Ruffian_210 : 50 (+ 31,= 10,- 9), 72.0 %
Spike1.2 : 50 (+ 16,= 14,- 20), 46.0 %
Naum : 50 (+ 23,= 14,- 13), 60.0 %


chessbase/noomen 25 opening test set
Level: Blitz 1/1
Hardware: Dual Intel(R) Core(TM)2 CPU T7400 @ 2.16GHz with 2,048 MB Memory
Operating system: Microsoft Windows XP Professional Service Pack 2 (Build 2600)

[Graham Banks]

Hi Joseph,

as a matter of interest it would be great if you could test other selectivity settings to see if you can duplicate Cock de Gorter's results that indicate selectivity 21 is likely best.

Regards, Graham.

[ozziejoe]

ok, graham, i'll give selectivity 25 a try. It does look like cm does not get as deep as the other engines.

best
J

[Kaj Soderberg]

Raising the selectivity level mostly works out positive with CM, although this looks like a lot.
On the other hand, blitz level does not say that much about chess playing quality.
Are you planning to do some testing on let's say 20/20 or 40/40 level?
That would say more about the chess playing capabilities of CM "sel21".
Anyway, thanks for the finding. It gives something to think about when fiddling with te parameters.

Best regards,

Kaj

[ozziejoe]

yes, i don't know how strong it is with long play, which is really the crucial question.There are so many parameters to be varied , though, and i have so few resources (one dual core to run at night and weekends), that i've got to try to narrow down the possbilities.


I vagualy remember something about an earlier version of rybka (and someone correct me if im wrong).Rybka had a selectivty parameter, and I remember vas suggesting a likely range of numbers. Later, he and testers discovered that the number needed to be much higher than expected, and outside the range....I don't know if my memory is right?


best
J

[Kaj Soderberg]

yes, i don't know how strong it is with long play, which is really the crucial question.There are so many parameters to be varied , though, and i have so few resources (one dual core to run at night and weekends), that i've got to try to narrow down the possbilities.


I vagualy remember something about an earlier version of rybka (and someone correct me if im wrong).Rybka had a selectivty parameter, and I remember vas suggesting a likely range of numbers. Later, he and testers discovered that the number needed to be much higher than expected, and outside the range....I don't know if my memory is right?


best
J

This "out of the range" intrests me as CM is concerned. The selectivity values you use, make me wonder at what kind of values the engine still reacts "normally", that is to say does just narrow the selection and move more swiftly through the plies. Any experiences?

My best,

Kaj

[ozziejoe]

I've gone as high as 25 selectivity, and cm did not seem to behave abnormally. It was winning about 49.5% after 60 games, about the same as sel =21


Raising the selectivity does seem to allow cm to get to greater depths (at least it seems so based on my nonscientific observations). Default CM seems to be getting consistently less deep than the other programs (even ruffian), so selectivity seems like a reasonable paramater to increase.

I've decided to look at sel =21 and king safety = 160. I watched CM get pawnstormed 5 straight times on the black slide of the caro kann, and I am thinking maybe it should not be so quick to castle into a pawn storm. We will see.

Can anybody tell me what it means to set opponents king safety to 160.?Does that mean cm will try harder for a kside attack?


other observations....., I agree with graham....cm does seem particularly human like in some ways. I love playing against its dummy downed personalities... It would be nice to increase its strength whilst keeping some of its interesting style.

best
J

[Tony Thomas]

After reading all this, I have decided to test with selectivity 21 as well.

[Dr.Wael Deeb]

After reading all this, I have decided to test with selectivity 21 as well.

A reasonable choice I assume,I'll catch the same bus

[Kaj Soderberg]

After reading all this, I have decided to test with selectivity 21 as well.

After reading all this promising stuff about extraordinary selectivity values, i would like to link my CM 11 to my Fritz 10 interface.
Is there a known step-by-step procedure for it, or even better, a turnkey solution?

Thanks for answers in advance.

Best regards,

Kaj