stephenbrooks.orgForumMuon1GeneralLinac900Ext10d2 different designs
Username: Password:
Search site:
Subscribe to thread via RSS
2010-09-29 11:57:31
As you can see from this plot: there are three different distinct designs that are increasing in % right now.  At 2500Mpts, at 1300Mpts and at 1500Mpts.

The big question is, which will win? 
2010-09-29 12:31:20
the best will be around 2600 Mpts
2010-10-01 14:00:32
Finger trouble..
[Edited by [TN]steinrar at 2010-10-01 14:03:09]
2010-10-01 14:01:34
My guess is 1500Mpts will win
2010-10-05 21:31:38
My guess is just below 1500Mpts.  It's increasing steadily in score and is now at 2.13! 
2010-11-04 12:52:08
Still betting in 2600 Higher Mpts designs allways prevail.

It will be the same with 6tc2 lattice.  Seems like 2000 Mpts designs hit the wall allready.  Now is turn on 2600 to get higher.  Actually it is kind of waste to concentrate on lower Mpts instead on higher.
2010-11-04 18:12:05
I guess you're right, but I also guess that the designs we are improving down at 1800 at the moment will improve the total % when we step up to 2650 ;-P

[DPC] Mr. Aldi
2010-11-04 20:54:07
I had some designs that had a lot more particles but lower yield.  Now, I'm trying with only high mpts results in .dat-file to create better results for these designs.  I'm running muon1 -/10 to speed this process up a little When designs improve, I'll do a normal scan

It seems that everybody is crunching on the the same accelerator with only minor changes in the Linac.  I guess Stephen also get lot of double or multiple times calculated configurations
[Edited by [DPC] Mr. Aldi at 2010-11-04 21:09:21]
[DPC] Mr. Aldi
2010-11-05 00:15:27
Heh heh

This is working already.  Soon I'm the one to post a configuration higher than 2.50 for 6tc2. Too bad you don't get extra Mpts for that

I use a muon1 -/10 to find interesting configurations.  As yields with /10 are typically about 0.2% lower, my current 2.38% could actually be 2.58% (but prolly is something just higher than 2.50
[Edited by [DPC] Mr. Aldi at 2010-11-05 00:15:49]
Stephen Brooks
2010-11-05 12:40:28
TOMAZ (2.585339%) seems to have beaten you to it.  I don't know if any manual intervention was involved or the client just did it on its own.
[DPC] Mr. Aldi
2010-11-05 13:22:54
Damnit!  Tomaz felt the heat already I guess.  Congrats Tomaz

All clients were crunching on the lower mpts.  I guess Tomaz selected the high mpts results for results.dat so that muon1 would only search for these kind of configurations.

Well, then I quit searching for higher yields as Tomaz result will now be taken by all the other clients for further search.  (I just got above 2.50 )
2010-11-05 13:27:06
As some time ago, when it got to 0.38 % I combined two designs.  Basic ingredient was Boinc's genome wich had highest yield among high Mpts designs:
2.456495 (12804.7 Mpts) [v4.44d] <Linac900Ext6tc2> #time=108393; by boinc (Boinc Wrapper)

Than I took than highest yield design (2.5..) and copy all db data to the boinc/s result.  So, the new monster design was born
2010-11-05 13:44:16
It sure seems, that high Mpts designs keep more particles for a longer time in the machine, but at some point something get wrong.  So, optimizing that point brings us high particle yield machine.
Iteresting test would be, if we all try to optimize just one or two high Mpts designs (and exchange result.dat file) for a week or two.  Stephen, can u arrange dat file in that way?  Or maybe yoyo from boinc ?

Candidates, for example would be:
1.752650 (2707.0 Mpts) [v4.44d] <Linac900Ext6tc2> #time=108439; by TOMAZ


0.405465 (3083.3 Mpts) [v4.44d] <Linac900Ext6tc2> #time=108385; by TOMAZ
[DPC] Mr. Aldi
2010-11-05 19:39:06
It would be possible to edit the sample-file to only contain those two or maybe better: make a 'new' lattice file which is actually the same scan but with a samples file that only contain the two configs above.  Then it would not interfere with the current results.dat and all the clients would pick this scan up automatically.

Exchanging results.dat would be a good thing to avoid dupes (if there are many dupes).  Otherwise a samples file (with a new lattice) should be enough?

tantalumrodr=006;tantalumrodz=243; First time I've seen a non-zero tantalumrodr
Stephen Brooks
2010-11-05 23:27:15
You can change the URL for sample files download (in config.txt) to point at another server, so constructing your own sample file sorted by other criteria is possible!
[DPC] Mr. Aldi
2010-11-06 12:41:26
I guess the 3000+ machines are never going to improve beyond 0.4x %. I ran your record on -/10 for a while but it only improved 0.03%. After a while it found lower mpts to be more interesting (and had to delete it to let it keep searching for the high Mpts).

If you look at some other optimalizations, you'll see it is not always the highest Mpts that prevail

Is a good example of that.
2010-11-06 16:48:45
The sample file which is send to the Boinc clients contains the 200 highest Mpts results and the sample file from Stephen concatenated as one sample file.  If you want to have it, just connect a Boinc client and download a Muon work unit.
You an see in this picture (the red line) , that the samplefile is generated nearlz every 8 hours.  But for the high Mpts result only from the Boinc results.
2010-11-06 17:20:37
I see some have picked up on my excessive Mpts, but lower yield result.  This was a stolen result from another lattice.  the high Mpts interested me as I was going to look to fine tune it by adding sections of other designs.  The results have so far been uninteresting, the best so far has been

1.893799 (10110.4 Mpts) [v4.44d] <Linac900Ext6tc2> {26685DB70003BFE3A2781626}

its getting more efficient.  The one thing that I hope is not a problem is the other lattice was incorporating the PhaseRotation.  So those parts are carrying over into this lattice designs.  I'm not sure how this may be affecting things.
From what I remember the phaserot cells were attempting to affect the energy of the particles in some way, but I cant remember what exact effect they had.

[DPC] Mr. Aldi
2010-11-06 18:55:05
waffleironhead, you see that your result actually evolved towards the 2000 Mpts region. 

PhaseRotation is doing nothing Compare:

0.430277 (3051.9 Mpts) [v4.44d] <Linac900Ext6tc2> {A72E75D069C3C...


0.430277 (3051.9 Mpts) [v4.44d] <Linac900Ext6tc2> {206BC6B164847C....

(I've shortened the hash so that you can't steel my points

Other results if somebody's interested (especially last one which is close to 2700 Mpts):

2.542096 (13261.8 Mpts) [v4.44d] <Linac900Ext6tc2>

2.509210 (2686.4 Mpts) [v4.44d] <Linac900Ext6tc2>
[Edited by [DPC] Mr. Aldi at 2010-11-06 18:56:40]
2010-11-06 20:30:13
Thanks for those results Mr.Aldi.  I hadnt got around to weeding out all of the PhaseRot specific parameters.  I'm glad that they are rendered useless by the current lattices.

Yes I did notice that the design had moved towards the 2k mpts area in its climb.  I'm still manually adjusting the design every so often, while letting it do its own thing.  I had forgotton how much I enjoyed these manual tweaks.
[DPC] Mr. Aldi
2010-11-06 21:12:17
Heh heh, currently I am also babysitting a client running @ 10 times speed to find interesting designs.

One client was constantly trying out the 3000 Mpts region, but it is still not past the 0,43. If there is room for improvement, muon1 quite quickly finds a better result.  From 0,43 the way to 2,60 is too long, so I guess ~2600 Mpts is the max.

The high speed client which found 2.509210% @ 2686.4 Mpts is now working on some improvement for the 10d2

I also tried the real manual tweaking (combining accelerators etc), but it didn't help.  I had a design for with a lot particles when it started but yet a low yield.  I now just trust Stephens trialtypes

btw, most probably PhaseRot does nothing for any result, but at least it doesn't for what I calculated.  I thought I also read somewhere on this forum about PhaseRot not being useful.
2010-11-08 09:56:57
[quote]You can change the URL for sample files download (in config.txt) to point at another server, so constructing your own sample file sorted by other criteria is possible![/quote]

Runesk has made a separate samplefile server that we use for our search for alternate designs for Linac900Ext10d2. It looks like we're just about to max out 1800Mpts.  Unless some interesting results show up the next couple of days, we'll move to 1900Mpts. 
2010-11-08 09:59:36
Stephen Brooks
2010-11-08 12:41:42
6tc2 has overtaken 10d2!  But the old optimisation has itself made a big leap.
2010-11-08 20:17:38
Stephen, does that mean that work done for 10d2 will be redundant?

2010-11-08 20:19:51
Btw, here is a scatter plot (from a an other tool I've made) containing all samples > 0% from my sample server.

Stephen Brooks
2010-11-09 00:08:03
--[Stephen, does that mean that work done for 10d2 will be redundant?]--

Not really, that optimisation has more linac sections so could end up higher in the end, it might just take a very long time to get there.  Plus its rate of increase has picked up in the last couple of months.  I'm leaving both of them up for now.
2010-11-11 19:46:43
I have figured out a simple rule for dat files.  If new best design have higher yield and Mpts than previous best, than it shoud be included.  Otherwise (higher yield but lower Mpts) it shoud be avoided because it leads only to local extrema. 

Yoyo, Runesk, Stephen are you willing to share results file manipulation software ?  It would be interesting to run own server for results.dat but can't prepare files "by hand".
Stephen Brooks
2010-11-12 12:28:54
Depends what programming language you want to write your sample-generating software in.  PHP (or maybe Perl) would be slightly easier than C because it involves manipulating text and arrays a lot.
2010-11-12 12:38:07
We're starting to close in on the current best design now.  We have 2.6% @ 2300Mpts in our alternate optimization. 

I still believe that we should focus on best muon % for samples in the starting phase of a new lattice, and then switch to Mpts when things start to stagnate.
2010-11-12 13:01:53
thomaz, you can get a copy of my solution. 

It's a Perl solution, using mod_perl.  It's not very robust, as it tends to get a corrupt database from time to time.

I'll try to prepare a package for you next week, if you want it.

You'll need an apache2, and a small number of CPAN modules (I think all is available in Ubuntu LTS).

These are the Ubuntu packages I use:

[Edited by runesk at 2010-11-12 13:02:08]
[DPC] Mr. Aldi
2010-11-15 23:32:10
Tomaz, I do not completely understand how you made that 3000+ Mpts design.

As some time ago, when it got to 0.38 % I combined two designs.  Basic ingredient was Boinc's genome wich had highest yield among high Mpts designs:

Than I took than highest yield design (2.5..) and copy all db data to the boinc/s result.  So, the new monster design was born

In the Boinc genome I don't see the phaserotcells (they don't do anything) (which are included in your 3000+ Mpts designs)?  How did they get into your new monster design?

Also: you say you combined a high yield high Mpts result with a high yield result.  How do you know that this would deliver a much higher Mpts design?

When we look at this optimalisation:,6

You can see that there even is a 3500 Mpts design!  How can we create such a thing for 6tc2?
2010-11-16 13:31:54
BTW; sending in a result now, that gives a better yield than the higher Mpts-results on 10d2.  Look out! 
[Edited by [TN]marvik at 2010-11-16 13:34:04]
[DPC] Mr. Aldi
2010-11-16 16:29:12
Excellent work.  I don't understand how all the other machines don't find higher results past the line accross the white parts @

<a href="">Yield vs. Mpts plot Linac900Ext10d2</a>

How many results are in your enhanced sample-file?  Should this also not be used for the entire project?  Many machines are crunching for nothing now...

For 10d2 we need a 3000 Mpts result
2010-11-16 17:22:45
Excellent find Marvik! 

I suspect the reason this happened is because there is too much focus on high Mpts in sample files too early in the simulation process.  That leads us to miss other potentially interesting designs.
[DPC] Mr. Aldi
2010-11-16 18:19:33
Maybe, but why didn't other clients not find this result while they did calculate a lot in the 2300–2350 Mpts region?
2010-11-16 18:36:22
@aldi: Perhaps because they first were missed by high-Mpts passing right by, and like opyrt says we have started lower than 2300 Mpts trying to get good designs and working slowly but steadily upwards using interesting samples.

I don't know the definate answer as to why it works.  But I'm sure there are more competent people here than me to answer this.  Perhaps there's been some luck involved too?

By the way, I was using only one sample in the samplefile when finding this result.  The coordinates was:,128

And this was the sample:
2.678461 (11807.5 Mpts) [v4.44d] <Linac900Ext10d2>

2010-11-16 19:02:32
I'll try to reformulate my thesis on why this can happen.

The main reason these optimization is not caught by the majority of clients simulating is due to a combination of the genetic/evolutionary algorithms and the focus on high mtps in the client. 

As various optimization paths are being discovered early in the project phase, the sample files being filled with high mtps results will lead to other paths having lower mpts at the moment, will not be prioritized.  The 12 genetic/evolutionary algorithms (not counting #gen=0 because it is "random" ) will all try to optimize the best result(s) known in results.dat and the samples. 

This approach could be described as "jump to conclusion" as it will neglect "interesting" paths at lower mpts.

The approach we have used to seed our samples, has been to locate "standalone" "high" score results at lower mpts.  By doing this for some time, we have gradually produced new interesting such results at higher mpts, hopefully carrying some of that optimization path.

As marvik have proven today, this approach has carried fruits. 

We've coordinated our work, using an other sample generator, built for this purpose. 

Opyrt should be recognized as the brains behind this approach, being the first among our team to spot these alternative, interesting paths. 

[Edited by runesk at 2010-11-16 19:17:06]
[DPC] Mr. Aldi
2010-11-16 23:08:38
Ok, thanks for your explanation.  That means you have to go through quite a lot of results and be lucky that this leads in the 'other path'.

I agree that it is necessary to improve low Mpts result first before you can improve the higher Mpts results.  I also found the current high for 6tc2 using the highest lower Mpts result with a lower very high Mpts result in the sample file (accidently though).

Probably you know which parameters to look for (tantalumrodz r for example)
[Edited by [DPC] Mr. Aldi at 2010-11-16 23:18:26]
[DPC] Mr. Aldi
2010-11-16 23:50:47
I get the feeling muon1 optimizes more in a horizontal way (at least for the current lattices).  For 10d2 there is no result higher than 2500 Mpts.  I'm sure that could go higher, but the clients haven't found anything there yet...

It could be ofcourse that potential high Mpts results first give a lower Mpts @ low yield and therefor are neglected.
Stephen Brooks
2010-11-17 17:08:31
Muon1 and the samplefile will choose results preferentially by muon % yield only (so the ones that tend to be at the far right of the plot).
2010-11-18 09:25:00
Stephen, I'm aware that your samples will choose % yield above Mpts, but I also know that yoyo generates samples using Mpts ad the main factor. 

As yoyo are the great producers of results, their production will lead the rest of us to their paths, like sheep. 

I'm not sure if there is a good way of automating sample generation that will provide Muon1 clients a godd set of samples to try to explore lower mpts and % yields.  Stephen, what will Muon1 do if it has high muon % yields in results.dat, but lower in the samples file?

Stephen Brooks
2010-11-18 14:15:07
--[what will Muon1 do if it has high muon % yields in results.dat, but lower in the samples file?]--

Muon1 merges the samples in as if they were part of results.dat and then chooses in a biased way by muon yield.  So in your case it will prefer to choose the higher results from results.dat (you could rename results.dat to stop it doing this).
[DPC] Mr. Aldi
2010-11-18 14:23:27
That means no client will ever automatically pick up:

3.301555 (3227.0 Mpts) [v4.44d] <Linac900Ext6tc2> #time=108801; by [DPC] Mr. Aldi

to look if this results can be improved.  As the current high is so much higher.
2010-11-18 16:07:45
[quote]That means no client will ever automatically pick up:
3.301555 (3227.0 Mpts) [v4.44d] <Linac900Ext6tc2> #time=108801; by [DPC] Mr. Aldi
to look if this results can be improved.  As the current high is so much higher.[/quote]

That is not correct.  The default sample files consist of a mixture of high %, high Mpts and random results.  So there is still a chance it will be picked up.

Yoyo's samples are 50% high Mpts and 50% high muon% (if I remember correctly), but I think Yoyo samples are only created from results received from Yoyo clients.
Stephen Brooks
2010-11-18 16:37:27
There is a small probability that Muon1 will choose to evolve from *any* result in its files, though of course the probabilities are a lot higher for the results with the highest muon yield (otherwise there would be no incentive to push the yield higher!)
2010-11-18 17:14:36
[quote]Yoyo's samples are 50% high Mpts and 50% high muon% (if I remember correctly), but I think Yoyo samples are only created from results received from Yoyo clients.[/quote]
: contact : - - -
E-mail: sbstrudel characterstephenbrooks.orgTwitter: stephenjbrooksMastodon: strudel charactersjbstrudel RSS feed

Site has had 25596459 accesses.