stephenbrooks.orgForumMuon1GeneralAdmin publishes super result
Username: Password:
Search site:
Subscribe to thread via RSS
gogomaus
2002-08-26 03:34:35
Hi Stephen,
found your short dialogue with Pascal in thread "progress twds v4.22". As I´m not able to link or shift it, just copy/pasted it hereto.
------------------------------------------------
quote Pascal:
Hey Stephen, congrats to a yield of 2.454 percent. 
What did you do to achieve such a high yield?
answer Stephen:
Pascal: yes, that was got by doing a rather simple thing: making all the solenoids as large, long and powerful as possible and making all the gaps as short as possible.  If this turns out to be the optimal configuration after all then this project will have been rather boring, but I'm hoping maybe we can find a different design with even higher yield than this.  If it does turn out to be "boring", version 5 will simulate the third and final section of this design too, and the optimisation for that ought to be non-trivial.
-------------------------------------------------

First a more technical question : Does "large, long....as possible" mean within the limits of the design parameters (000-999) ?  (I do believe, but f.e. rechenkraft forum considered it to be out of limits and therefore invalid).

By far more interesting : What´s your intention to publish it just now (without any further comment) ???
Everyone would have accepted the announcement to work on a big playground just to check the power of genetic algorithms on a more complex design.
Now it _looks_ at least, you´re just kidding participants regardless of their more or less serious intention to support a scientific design simulation.

Ciao, Wolfgang
Pascal
2002-08-26 05:20:45
..making all the solenoids as large, long and powerful as possible and making all the gaps as short as possible..

For me it seems that some parameters are within the possible ranges, because he said that.  It's quite an interesting question to know wether he took the biggest-as-possible ranges for the simulation.

Otherwise, although he found one solution, it might not be the one with the best yield at all, might it? 
___________________________
Member of www.rechenkraft.net - German Website about Distributed Computing Projects

1: Athlon TB-C, 1.2 GC/s, 256 MB DDR-RAM, Erazor x², ADSL-Flatrate, NIC Intel, Win 98 SE Mainboard MSI-6380 Rev.  1
2: Pentium III, 600 MC/s, 256 MB RAM, NIC Intel, Win 98 SE
Stephen Brooks
2002-08-26 12:50:17
quote:
Originally posted by gogomaus:
First a more technical question : Does "large, long....as possible" mean within the limits of the design parameters (000-999) ?  (I do believe, but f.e. rechenkraft forum considered it to be out of limits and therefore invalid).


Nope, it's within limits but only just: all except one of the values are either 000 or 999.

quote:
By far more interesting : What´s your intention to publish it just now (without any further comment) ???


Because I hadn't actually tried inputting _those_ particular parameters before, so I did.  I think I'll leave my own results out of the best250 from now on, though, because I want to see if the genetic algorithm independantly comes up with this design as the best (in which case it will have been a rather uninteresting optimisation), or a different one (in which case it will have discovered something new).

quote:
Everyone would have accepted the announcement to work on a big playground just to check the power of genetic algorithms on a more complex design.


This is certainly what we have done.  I have talked with the overall designer of this accelerator and it turns out that all the "narrowing down" etc. that I was advised hard-code into version 4.1 was partly to do with the _third_ component of the design (the linac), which won't be simulated until version 5. I didn't know this before and thought that just simulating these two components together would produce a non-trivial optimised design.  Now I'm not quite so sure if this will happen until the linac is added.  So there are two possibilities:
  • The 4.2x optimisation turns out to converge to my extremal design, in which case it will have been a bit 'boring', but still gives important information about the design - i.e. that for transmission through these two components, we simply need as much focussing as possible (also interestingly, my new design used solenoids of all the same polarity, whereas all v4.1x designs had them alternating in some way, so it would show polarity is not important either). 
  • The 4.2x optimisation finds a different design with an equal or even higher muon yield.  In this case we will have discovered something very new, in a large parameter space.
Either way, getting convergence in this space is something of a success.  If the first case is true, then the 2.45% result will at least allow me to tell how good the algorithm is in terms of approaching the best result.
If people are discouraged by this finding, you can stop running the program entirely, or maybe just resume in version 5 when the full design is being tested.  But certainly one real success so far has been getting this DC-platform set up in the first place, and now with a demonstrably powerful optimisation engine.  Some of the upgrades to versions 4.22 and 5 are not aimed at the DC-users, but at the other "users" - the scientists - who want to input their designs to be tested.  One thing I've been working on is a generalised file format for specifying designs and their optimisation ranges for particle accelerators, so when this is in place it could be possible that we will be optimising many sorts of accelerators.


"As every 11-year-old kid knows, if you concentrate enough Van-der-Graff generators and expensive special effects in one place, you create a spiral space-time whirly thing, AND an interesting plotline"
gogomaus
2002-08-26 17:58:18
quote:
Originally posted by Stephen Brooks:

Because I hadn't actually tried inputting _those_ particular parameters before, so I did.  I think I'll leave my own results out of the best250 from now on, though, because I want to see if the genetic algorithm independantly comes up with this design as the best (in which case it will have been a rather uninteresting optimisation), or a different one (in which case it will have discovered something new).


As a scientist I would have expected to do some obvious extrem design test in advance before getting DC users for the hunt.  It would have been nice to know safely there is at least one better result than under v4.1x. Your arguments for both alternatives (getting that one confirmed or finding even better ones) would have been valid and more informative, if announced while starting v4.2.

O.k., nobody is perfect and perhaps I do belong to a critical minority (as nobody else wondered about your premium result).

My disappointment is rather limited, as I received a lot of semi-intellectual entertainment and many competent answers from your side since entering this project as a long distance member.

Adding the linac and/or testing other accelerator concepts sounds very interesting as a future outlook.
Bye
Stephen Brooks
2002-08-27 05:14:51
quote:
Originally posted by gogomaus:
As a scientist I would have expected to do some obvious extrem design test in advance before getting DC users for the hunt.  It would have been nice to know safely there is at least one better result than under v4.1x. Your arguments for both alternatives (getting that one confirmed or finding even better ones) would have been valid and more informative, if announced while starting v4.2.


The trouble was, that there was no such extremal design under 4.1x since the parameter space was more limited and also the "best" 4.1x result appeared to be quite far from this extreme (S2 was high, but S3 and S4 were considerably below-maximum).  I had also been led to believe that such an extreme design would not be any good becuase the original prototype design (by someone else) had quite carefully narrowed down the solenoids and varied their strengths.
So it was only later on that I thought that I should test this design "just in case", and that's why it appeared now instead of at the beginning of the optimisation.  After I'd tested this, I reported my finding to the original designer of the section (i.e. 'why do we need narrowing etc. when this simulation suggests we just need as big as possible?') and only then did he tell me that these features of the decay channel were to do with the linac that I have not yet simulated.


"As every 11-year-old kid knows, if you concentrate enough Van-der-Graff generators and expensive special effects in one place, you create a spiral space-time whirly thing, AND an interesting plotline"
gogomaus
2002-08-28 04:33:03
quote:
Originally posted by Stephen Brooks:
I think I'll leave my own results out of the best250 from now on, though, because I want to see if the genetic algorithm independantly comes up with this design as the best (in which case it will have been a rather uninteresting optimisation), or a different one (in which case it will have discovered something new).


Just another change of mind ?  wink

Your actual Best 250 result.dat does include your "extreme" results, so at the given advantage in yield-%, it will lead everyone who appends it directly to this top edge of design.
Was it an error or did you like to ring in the final round for this (version) experiment ?
[SG] MiniCruncher
2002-08-28 04:53:14
And what shall we do now ??????

Should we delete this "extreme" result ??

And since this new best250 I habe a "2.262829 (61086 particles)" result.  Is this ok or is it wrong because i think it belongs to your "extreme" result ??

Ciao Rudi
AySz88
2002-08-28 07:20:18
I've been thinking...

If the force tapers off the closer to the center of the beam you get, then for the first few solinoids, largest would not be optimum.  Some particles are doomed to crash into a solinoid because of the angle at which they come out of the first solinoid (solinoid prime?  big grin ).  With only largest solinoids, some angles that would be a close call and barely curve into the beam in the best optimization would instead recieve less force than the ones that are doomed to crash into a solinoid, and the close-call particles would not curve enough.  They fly out instead of staying in the beam.

[This message was edited by AySz88 on 2002-Aug-28 at 20:08.]
Stephen Brooks
2002-08-28 08:46:58
Yup, changed my mind about letting the high result in the best250.

More interestingly the stats appear to have stopped at around 10am today so I'll be wanting to sort those out when I get home.


"As every 11-year-old kid knows, if you concentrate enough Van-der-Graff generators and expensive special effects in one place, you create a spiral space-time whirly thing, AND an interesting plotline"
Stephen Brooks
2002-09-03 09:18:44
I did a few runs of my "extreme" result on this computer and got four percentages:

2.368442 (64078 particles) [v4.22] {89467834}
2.389959 (64196 particles) [v4.22] {0A012783}
2.381033 (64272 particles) [v4.22] {836218F2}
2.405463 (64073 particles) [v4.22] {DEDA9B1B}

Using the maximum likelihood estimation from these (assuming a normal distribution), we see that this configuration will produce results with mean 2.3862% and standard deviation 0.01557%.
With these figures, the probability of this design scoring >2.4% is 18.8% (quite high).
The probability of it scoring over 2.45% is 0.0021%: one in 48`000 runs, so we probably would see this deviation about once in all the calculations that have been done since I added the result.  Pascal mentioned a 2.454% result above, but I think that was a while after I'd added this result to my own files, so it had started to mutate by then.
The probability of it scoring over 2.5% is 1 in several trillion, which is essentially impossible.  And we've got all the way to >2.7%, so I think it's very unlikely that it's still this result giving the high scores.


"As every 11-year-old kid knows, if you concentrate enough Van-der-Graff generators and expensive special effects in one place, you create a spiral space-time whirly thing, AND an interesting plotline"
Pascal
2002-09-03 10:20:58
Hm, Stephen, now I have a little question more:

may we use the results of version 4.21b in version 4.22?

___________________________
Member of www.rechenkraft.net - German Website about Distributed Computing Projects

1: Athlon TB-C, 1.2 GC/s, 256 MB DDR-RAM, Erazor x², ADSL-Flatrate, NIC Intel, Win 98 SE Mainboard MSI-6380 Rev.  1
2: Pentium III, 600 MC/s, 256 MB RAM, NIC Intel, Win 98 SE
Stephen Brooks
2002-09-03 10:56:47
Hmm Pascal, I have a little answer here: 'yes'

Well actually, 'no' because the checksums are different.  I'll have to put in something that switches to the old form of checksumming for versions below 4.22, so it doesn't just ignore the previous results.  The designs are interchangable though.
gogomaus
2002-09-03 12:37:01
Hi Stephen (re your probability chat),
arguments concerning your _initial_ design are o.k. (ignoring that 4 is a rather low n for a statistical base), but only valid for this one.  How many of new high score results do really comply exactly to it ?
I assume most (if not all) of newer incoming results are mutations/variations of your original design, so they are not really independent ones.  (Like the 2.454% example mentioned above, which was the first "super result" I saw in the stats, which of course, I did not supervise night and day.)
Stephen Brooks
2002-09-03 13:53:44
The current best designs probably combine some features from both the super-result and the original ones we had already, or at least are fairly-mutated versions of the super-result.  So we made progress beyond it regardless.  I don't see anything wrong with that.


"As every 11-year-old kid knows, if you concentrate enough Van-der-Graff generators and expensive special effects in one place, you create a spiral space-time whirly thing, AND an interesting plotline"
pben
2002-09-03 20:50:25
The Advanced option in the muon1 viewer is a neet tool.  It is interesting to see a vast plane execpt for a mountian piled up in in the corner or along an axis.  I have no idea what the variables mean but clicking on the variables in the first column generaly show a small range of good results even in the 4+ megs of my results.dat.  I hope that you include a similar tool in future versions.  The 3D plus color is a great feature.

Thanks
stoyan@bravo
2002-09-04 00:20:13
About the probability chat. 
Stephen, I am not quite sure that the normal distribution is fine since the support of the distribution is the whole real line so there is a non-zero probability that the result is negative which is nonsense.  I would suggest fitting a beta distribution since:
- the support is the interval [0, 1], 0 corresponding to 0% and 1 corresponding to 100%
- the beta family contains a variety of bell-shaped curves
I would do the fitting myself but I do not have a good sample.
Stephen Brooks
2002-09-04 01:58:50
The beta distribution is indeed the correct one to use in this case, but the formulae for MLEs with beta functions are fairly messy, and my calculator just happened to have the normal ones programmed in.  The normal distribution seems to be OK in this case since the probability you derive of getting a less-than-zero result is something like 10^-120 (maybe less).  This is a common approximation to make.  What I'm less happy about is the small sample-size so I might run some more trials of that design at some stage.


"As every 11-year-old kid knows, if you concentrate enough Van-der-Graff generators and expensive special effects in one place, you create a spiral space-time whirly thing, AND an interesting plotline"
Stephen Brooks
2002-09-04 02:45:14
Hmm I just inspected my animated results-histogram and got some better evidence: the initial "hump" formed by the extreme-result appears to be shifting rightwards (to higher %ages) as the optimisation progresses, which suggests a real improvement rather than just variation around a point.

cool The optimisation appears to be doing well - I didn't really expect yields of 2.8%, although I hoped it would do better than I expected.


"As every 11-year-old kid knows, if you concentrate enough Van-der-Graff generators and expensive special effects in one place, you create a spiral space-time whirly thing, AND an interesting plotline"
stoyan@bravo
2002-09-04 11:29:17
(Probability chat)Some additional remarks. 
Yes, Stephen, you are right for the MLE but since all moments of the beta distribution exist,
we can try the method of moments for the parameter estimation.  The formulae are very simple and about 10 000 observations should be enough. 
Here we also have another statistical issue.  An important assumption is the independence of observations.
Obviously in our case the observations are dependent since the program learns from the results achieved so far. 
So the probability of getting a new record should be not less than the number that we compute from the fitted distribution (we get a lower bound). 
And for the fitting, we should use the results based on one and the same result250 file. 
I will be very glad if I can be of help in this.
Stephen Brooks
2002-09-04 12:10:06
I was only intending to fit the distribution to runs produced from the _same_ design.  The distribution of results created by the whole project is much more complex - time-dependent, probably resembling some distribution about an upwards random-walk, plus another distribution for the 25% of random results, plus another scale of random walk for the best250 releases, etc. It's not really worth the effort!


"As every 11-year-old kid knows, if you concentrate enough Van-der-Graff generators and expensive special effects in one place, you create a spiral space-time whirly thing, AND an interesting plotline"
Euler
2002-09-05 04:59:19
Stephen,

now that you published a "top250", I'm wondering, whether it's possible to provide
a "complete" set of _all_ results ?

Euler
Stephen Brooks
2002-09-05 06:28:48
That would be about 350MB in size (40MB compressed).  It would take rather a long time to transfer to you and would probably not fit inside my web account either big grin.
MaFi
2002-09-05 09:12:31
But what do you think about a best1000 or best5000?
I think that would be quite interesting.  cool
Markus
Euler
2002-09-05 12:19:34
quote:
That would be about 350MB in size (40MB compressed).  It would take rather a long time to transfer to you and would probably not fit inside my web account either .



If that's a problem, I'd be happy, to mirror the file for you.

Euler
Stephen Brooks
2002-09-06 02:23:44
I can try putting up best1000 and best5000 files too, for those who don't mind a larger download.


"As every 11-year-old kid knows, if you concentrate enough Van-der-Graff generators and expensive special effects in one place, you create a spiral space-time whirly thing, AND an interesting plotline"
Ben Whitten
2002-09-09 07:11:18
Well I think you only need to include the results that are obvously a result of that "extreme" result thing you made cause the other results are begining to look like insignificant "ants" of numbers big grin and you probs dont want to clutter up your web site with extra crap that isn't going to be used....... are you?
Stephen Brooks
2002-09-09 11:48:59
They're all mixed together now so I'll just put up whatever's best.  When we go above 3% another best250 file will arrive (and also best1k, best5k for obsessives).
Ben Whitten
2002-09-10 09:22:27
Well I am takin a backseat on this project because I see all you with ur high end systems (50 computer's in one case I've seen) and my crapy 600Mhz takes to damn long to do one simulation (like 2 days on and off, like a normal computer) so I'm going to wait till the manual thing comes out, probs will be slower but my computer just cant hack what I want it to do.  frown
[ARS]odessit
2002-09-10 10:44:02
Why quit?
1st - Every machine counts!  Regardless of speed as long as it makes SOMETHING
2nd - run your machine 24/7 wink
3rd - build another faster machine and see #1 big grin

I am running this project on my laptop (besides my home machine), provided it is 200 MHz faster than your machine but with erradic crunch times it's still making some progress.  I wish I would have your 600 MHz, I'll take it and will keep on running Muon1 razz
My 2 roubles (0.00002 cents)

Intel P3-M 866 MHz
Duron 1000 MHz
XP 1600+ @ 2000+
Ben Whitten
2002-09-10 13:13:03
ya see I cant run my comp all night cause it's in my bros room temorerily, and once it's in the other room it's right nxt to my mum and dads room (they dont like the fan noise) but I could be gettin a 2.53Ghz after my exams big grin woooopty dooooooo
AySz88
2002-09-10 19:06:22
Ben: get a 2.8 GHz; they're out, you know.  smile
Bluumi [SwissTeam.NET]
2002-09-10 23:26:49
@Ben Whitten
I don't use the 50PC again ... better?  big grin
In the Moment, i can't use it again... i use only my home pc... and now you have a chance

But you'r right, the bigger Muon-Results are much slower... and with a slower pc is need many hour to simulate one... roll eyes
Stephen Brooks
2002-09-11 01:27:03
quote:
Originally posted by AySz88:
Ben: get a 2.8 GHz; they're out, you know.  smile


Well that's what I've got, but I'm "cheating" because it's really a dual AMD1600+ (2*1.4GHz = 2.8) big grin
It's probably wise not to get the _Very_ fastest machine out unless you have serious money to burn.  There seems to be a price-curve at any one time that goes up gradually and then zooms right up at the top 10% of speeds available.  If you compare this with how the curve evolves in time as well, you find that a more moderate computer will only lose say 10% of its value in the next 3 months, whereas one of the really fast ones might lose 30% of its value...
: contact : - - -
E-mail: sbstrudel characterstephenbrooks.org
Yahoo: scrutney_mallard
Jabber: stephenbrooksstrudel characterjabber.org
Twitter: stephenjbrooks

Site has had 15715368 accesses.