stephenbrooks.orgForumMuon1GeneralWhy is Linac88MHz900MeV1 so prone to duplicates?
Username: Password:
Search site:
Subscribe to thread via RSS
[TA]JonB
2008-04-07 15:31:03
Stephen (or anyone else), I've noticed many times were Linac88MHz900MeV1 values end up with 4 or 6 or now 10 duplicate results.

I've never seen it to that degree in any optimization before.  Can you think of any good reasons?  Is it a good thing, bad thing or who knows??

Thanks, JonB
Stephen Brooks
2008-04-09 11:30:55
The lattices ending in "1" have fewer variables than the other lattices, so there will be fewer configurations to search.  It's even possible this one might have found the optimum and converged, though more likely it's just got somewhere from which all the other designs it tries are downhill.
Stephen Brooks
2008-04-18 11:59:23
I think it converged!  Of course there's no way to know for sure other than an impossible exhaustive search of the design space, but the Mpts/yield distribution looked like it had climbed into a corner with no way out or up.

I'm now focussing the whole project on the Linac88MHz900MeV6 lattice, see if that can 'get around' the impasse.  The other PhaseRotLinacs were interesting but didn't accelerate the beam enough for those high 7% yields to really translate into something the accelerator can use.
tomaz
2008-04-18 14:57:23
It's funny.  Most results have linaccells=000, so there is no effect of linear accelerator.  Problem is, that designs with low linaccels converge faster.  If acceleration is problem, why you didn't make dat files only with "enough" linaccels ?  I just got 6.895 % muon result with linaccells=150 an I think it can go further.  I'll start to play with higher linaccels values.  Do you have a hint which value can give enough acceleration ?

I also have a bit of the topic idea: why compute all designs to the end ?  If at some point during simulation becomes clear that design isn't good (lower than highest in dat file or something like that) computation should be abandoned and new one strated.  It would save computation time.  To keep gene pool wide enough every 10th or so result should be computed to the end...
Stephen Brooks
2008-04-18 16:39:25
You are probably talking about PhaseRotLinac, in which case the optimiser is using the RF cavities in the phase rotation section to do the acceleration as well as the phase rotation (which is possible).

In the 900MeV linacs there is no way the beam can achieve the goal energy range without actually using the linac and this means linaccells must be quite high.  In fact, these lattices seem to have the decaycells set to 000, indicating a decay channel of the minimum allowed length!

--[I also have a bit of the topic idea: why compute all designs to the end ?  If at some point during simulation becomes clear that design isn't good (lower than highest in dat file or something like that) computation should be abandoned and new one strated.]--

1. Having lower designs fully calculated will give the optimiser a better idea of the trends surrounding a certain area.  If a calculation was aborted, there will be no muon score value to go with that genome, so from the point of view of the optimiser "learning", it would be wasted.

2. The designs that are not good and lose particles early also simulate faster (few Mpts) because there are fewer particles, so we are not really losing a huge amount of computing time.  The designs that are not good but only lose particles at the end (high Mpts) are difficult to identify until the particles are actually lost.
: contact : - - -
E-mail: sbstrudel characterstephenbrooks.orgTwitter: stephenjbrooksMastodon: strudel charactersjbstrudel charactermstdn.io

Site has had 17077828 accesses.