Statistics and scoring

[gd_barnes]

Originally Posted by jasong

I can't help but note that after the Twin Prime Search primes fall off the list a few months from now, a person could find themselves in the position of actively crunching the project, maybe even increasing in rank, and yet getting lower and lower scores.

Maybe we should archive a list, starting now, of the digit length of the prime at 5,000th place every first of the month. If someone gets discouraged, they have the option of making new comparisons based on a non-moving target, the value of the 5,000th placed prime in a given month.

Is it possible to simply calculate a simple normalization value? (Not sure if normalization is the right term. )

IMHO, just leave the 'normalization number' at n=333333. That way, everyone is normalized the same way at all times. Sure computers will be faster in the future and finding an n=350K prime will be much easier in 5 years than today but my thinking is that you have to be actively searching with us or your rank will slowly drop, even if your score does not.

I don't see a point in changing this normalization number based on the current 5000th place prime. Everyone's score will adjust downwards by the same percentage so it just confuses things to change it. That way people are never dropping in score. They just drop in rank if they aren't actively searching for us.

What do you guys think?


Gary

[Flatlander]

Originally Posted by gd_barnes

IMHO, just leave the 'normalization number' at n=333333.
...
What do you guys think?

Gary

Yes, keep things simple for my sake!
(After all, isn't GIMPS still using 'P90 years'?)

[kar_bon]

Originally Posted by Flatlander

Yes, keep things simple for my sake!
(After all, isn't GIMPS still using 'P90 years'?)

GIMPS? yo, but for a single n at very high values, compared to our mini-n.
they perform ECM on values at our level, look here http://v5www.mersenne.org/ to the new V5 test-server.
so think of our ranges: about 2300 k/n-pairs and for every one this had to be calculated.
i don't have any automatism (yet) behind all scoring. i'm thinking of php-scripts but that's far from now.

so this scoring is ok and i can do that only with Ex*el.

[gd_barnes]

Originally Posted by kar_bon

GIMPS? yo, but for a single n at very high values, compared to our mini-n.
they perform ECM on values at our level, look here http://v5www.mersenne.org/ to the new V5 test-server.
so think of our ranges: about 2300 k/n-pairs and for every one this had to be calculated.
i don't have any automatism (yet) behind all scoring. i'm thinking of php-scripts but that's far from now.

so this scoring is ok and i can do that only with Ex*el.

Karsten,

Are you OK with leaving the 'normalization' number at n=333333 even after they are no longer in top-5000? That would be easiest for the manual system that we now have.


Gary

[kar_bon]

Originally Posted by gd_barnes

Karsten,

Are you OK with leaving the 'normalization' number at n=333333 even after they are no longer in top-5000? That would be easiest for the manual system that we now have.

Gary

i'm ok with that. so or so, the score is compareable, only the value would be different!

[gd_barnes]

Karsten,

Another challenge for you on scoring...

How shall scoring be done on ranges for LLRnet testing?

For primes, obviously it's the same as before.

But for ranges tested?...

Here's what I will suggest: Get the # of candidates tested for each individual in the LLRnet range. Then divide it by the average file size of an n=200 file, which will be ~2300, and use that multiplier to apply to the usual range computation. Example:

a = avg. # of candidates in n=200 file
c = normal computation for an n=200 range
n = # of candidates tested in LLRnet range for a particular person

LLRnet range scoring for a particual person:

c * n / a

If you agree this is a good way to do it, I'll get you an exact average on the # of candidates in an n=200 file.

BUT...what happens for ALL range scoring (not only LLRnet scoring) when the posted file-ranges start becoming smaller as a result of testing higher n? I'll let you 'chew' on that one for a while!


Gary

[mdettweiler]

Originally Posted by gd_barnes

Karsten,

Another challenge for you on scoring...

How shall scoring be done on ranges for LLRnet testing?

For primes, obviously it's the same as before.

But for ranges tested?...

Here's what I will suggest: Get the # of candidates tested for each individual in the LLRnet range. Then divide it by the average file size of an n=200 file, which will be ~2300, and use that multiplier to apply to the usual range computation. Example:

a = avg. # of candidates in n=200 file
c = normal computation for an n=200 range
n = # of candidates tested in LLRnet range for a particular person

LLRnet range scoring for a particual person:

c * n / a

If you agree this is a good way to do it, I'll get you an exact average on the # of candidates in an n=200 file.

BUT...what happens for ALL range scoring (not only LLRnet scoring) when the posted file-ranges start becoming smaller as a result of testing higher n? I'll let you 'chew' on that one for a while!


Gary

Hmm...that's sort of what I was thinking, too, though obviously, as you mentioned, there are problems. Another possibility: We could just score LLRnet as a single user in the main stats, then people could look into the separate LLRnet stats to see how LLRnet's stats divide up into individual users.

[kar_bon]

yes anon, these were my first thoughts about LLRnet too:

- primes found goes to individual like 'normal' found prime

- ranges will score for 'member' LLRnet.

it's the easiest way to do the scoring, because there's no additional calculation needed.

any others opinions?

PS:
to make an extra table for LLRnet individual scoring i need the result-files and the scoring takes time to be shown.

[mdettweiler]

Originally Posted by kar_bon

yes anon, these were my first thoughts about LLRnet too:

- primes found goes to individual like 'normal' found prime

- ranges will score for 'member' LLRnet.

it's the easiest way to do the scoring, because there's no additional calculation needed.

any others opinions?

PS:
to make an extra table for LLRnet individual scoring i need the result-files and the scoring takes time to be shown.

Actually, as for the LLRnet individual scoring, I've already got that in the "LLRnet Stats" thread--so you don't have to worry about that.

[em99010pepe]

Karsten,

Please add LLRNet stats to your stats page. Thanks.

[kar_bon]

Originally Posted by em99010pepe

Karsten,

Please add LLRNet stats to your stats page. Thanks.

so i need completed ranges for LLRnet!