stephenbrooks.orgForumMuon1GeneralYields no longer rising, where to from here?
Username: Password:
Search site:
Subscribe to thread via RSS
John Kitchen
2003-05-20 07:40:02
Seems to me that I am stuck at a max yield of 13.665393, and that's probably an overstatement, my best result from multiple runs (9) is 13.633587.

If this doesn't change over the next day(s), then I wonder if I am "in a box", with a maybe good design, but unable to get out and explore other potentials?

I'm reluctant to just start again from zero, in the hope that I can get lucky and find another path to a better design, or can I rely on the client to extract me from this local area, and get me into a better zone?

Maybe I need to completely re-seed my results.dat file?

Ideas, people?  Anybody else "stuck"?

John
Stephen Brooks
2003-05-20 08:00:31
We might have managed to hit the maximum already.  I'll release a best100 or 250 file some time soon, to see if that gives people a boost.  One thing I've noticed is that these optimisations tend to go by pretty _fast_ with the number of people running this project, even though we're optimising a moderately large number of variables.

Don't worry about wasted effort though - at least not in the long term - because the neutrino factory design study is getting into gear.  Can't give details just yet, but the main study phase will probably start in 2004, with an increase in interest before that.  Anyway, I plan to use the Muon network to optimise even more challenging things in future - more difficult individual simulations + more variables to optimise.  The fact we're making mincemeat of the current optimisations in a matter of weeks is nothing but encouraging.

Oh - something you could do is manually go through the genome and look for parameters that have values near the ends of the ranges, like 017 or 994. You could set those to the nearest extreme values like 000 and 999 to see if that gets the last bit out of the yield.

Today's weather in %region is Sunny/(null), max.  temperature #NAN°C
px3
2003-05-20 09:34:34
quote:
Originally posted by John Kitchen:
Ideas, people?  Anybody else "stuck"?

John


Hi John,

we decided to restart from scratch, because i'm not really sure if our results are valid and the other reason: we can't get higher than 12.9xx. 

Regards,

PX3
[DPC]Stephan202
2003-05-20 11:27:14
Currently the yields in the DPC TopNNN file are also climbing very slow.  Highest at this moment is 13.374989.

Edit: a result with a yield of 13.525438 has just been uploaded Smile.

---
Dutch Power Cow.
MOOH!

[This message was edited by [DPC]Stephan202 on 2003-May-20 at 19:41.]
John Kitchen
2003-05-20 12:32:00
quote:
Originally posted by [DPC]Stephan202:
Currently the yields in the DPC TopNNN file are also climbing _very_ slow.  Highest at this moment is 13.374989.

Edit: a result with a yield of 13.525438 has just been uploaded Smile.

---
Dutch Power Cow.
MOOH!

[This message was edited by [DPC]Stephan202 on 2003-May-20 at 19:41.]


Check in your results.dat to see if the new high score is a single calculation or a "#runs=nn" multiple run calculation.  I was finding that the only way I was getting "better" yields was with single calculations, and if you do enough of these, you eventually get "lucky" because of the limited number of particles simulated.  But the design is not necessarily "good", and is probably no better than the lower scored ones.

Personally, I only regard the results of my multiple calculations as valid, and I am using 7 and 9 repeats depending on the speed of my machines.

Based on ignoring the results of single calculations, I am "stuck", and not seeing any valid new high scores.  Frown

I think px3 has a good idea, and that is to start again, and hope that you don't find the same design.  We have time on our side, so even if it takes several weeks to get back up to high scores, it doesn't matter.

And I think that this time, I will keep my two principal machines from seeing each other's results.dat in the hope that they end up in different "places".

My personal goal is create the highest yield, and not the highest number of particles processed.  And I want that yield to be valid and repeatable.  Science rules!  Big Grin
[DPC]Stephan202
2003-05-20 12:46:08
The top three in our topNNN file:
tantalumrodz=558;tantalumrodr=027;s1l=994;s1f=998;d1l=000;s2l=999;...s26r=990;s26f=967;#runs=005;
13.525438 (268.1 Mpts) [v4.31b] {2E8DA8E5} <SolenoidsOnly>
tantalumrodz=559;tantalumrodr=027;s1l=994;s1f=998;d1l=000;s2l=999;...s26r=990;s26f=967;#runs=005;
13.438899 (530.9 Mpts) [v4.31b] {220BDABB} <SolenoidsOnly>
tantalumrodz=558;tantalumrodr=025;s1l=995;s1f=998;d1l=000;s2l=981;...s26r=990;s26f=966;#runs=005;
13.374989 (1343.2 Mpts) [v4.31c] {A43330E2} <SolenoidsOnly>

As you can see, they have all three been rechecked 5 times.

Most likely we (DPC) will open our results up- and downloader to the public by tomorrow night.  We'll then see if this will improve the results within a week or so (more people calculating on the top results --> higher chance of a positive mutation).

If this does not improve the yield percentage, it might indeed be an idea to start all over again.

---
Dutch Power Cow.
MOOH!
John Kitchen
2003-05-20 13:18:26
Are you sure about these data?  5 or more runs will just about always give a "Mpts" value >1000 (typically about 1300).  Your third one looks typical, but the first two look strange.  I just checked 240 of mine and only one was Mpts<1000, and it was 502.8 Mpts.
[DPC]Stephan202
2003-05-20 13:44:55
You got a serious point there.

We just decided to open up the database.  Take a look for yourself at http://stephan202.qik.nl/

You may rerun them as much as you like Smile
Oh, and add your own results please Wink

---
Dutch Power Cow.
MOOH!
Stephen Brooks
2003-05-20 15:59:26
quote:
Originally posted by John Kitchen:
Are you sure about these data?  5 or more runs will just about always give a "Mpts" value >1000 (typically about 1300).  Your third one looks typical, but the first two look strange.  I just checked 240 of mine and only one was Mpts<1000, and it was 502.8 Mpts.


Yeah.  Note the version-stamp.  The top two are results of the "design jumps out of queue" bug in v4.31b that is now fixed.  I'd suggest look at the ratio of Mpts-to-yield on the pre-v4.31c results, and eliminate any stragglers below what v4.31c makes.

Today's weather in %region is Sunny/(null), max.  temperature #NAN°C
: contact : - - -
E-mail: sbstrudel characterstephenbrooks.orgTwitter: stephenjbrooksMastodon: strudel charactersjbstrudel charactermstdn.io RSS feed

Site has had 25161411 accesses.