Stephen Brooks 2014-01-06 19:06:21 | Looks like Linac900Ext8Xc2 isn't showing any more yield gains, although it's been a very bumpy optimisation over the past year with somewhere between 3 and 6 "break out" events where an apparent convergence plateau was replaced by a rapid rise. It also has the distinction of being the highest-yielding 900MeV linac we've designed under v4.45+, i.e. since the yield noise bug was removed. I've replaced it by 5Xc2, which ought to be quicker to optimise because of fewer parameters. I think I'll go all the way down 4,3,2,1 in future to have a proper comparison of the effect of different numbers of blocks in the linac. It also looks like _nosample is about maxed out, I may replace that in due course. Specifically, 8Xc2 reached a yield of 4.283614% muons per pion, ahead of 7Xc2's 4.142452%. Development of Muon1 for new accelerator applications is actually pretty active right now, though I haven't got a multiparticle problem that needs the full power of the distributed network just yet, so we continue with the neutrino factory here and I'm using it a lot just on my workstation for other accelerators. However, multiparticle tracking is definitely going to be needed for the other projects at a later stage. |
Stephen Brooks 2014-01-06 21:27:33 | First results on the new lattice: 1. [OcUK]Phil 53 1917.6 (1.00x) -0.336286 0 |
Dave Peachey 2014-01-07 18:22:06 | Stephen, when you decide to remove the current "_nosample" optimisation, would you replace it with a different "_nosample" one or not bother with that approach again? I ask because I rather enjoyed having something that relied on the individual efforts of each individual cruncher (and to which one could, as I chose to do, dedicate all one's efforts) rather than constantly referencing a single "best of" parameter set; or has this particular experiment (the concept not the optimisation set) not yield a suffuciently different set of results to warrant that? |
Stephen Brooks 2014-01-07 21:55:48 | Yes, the _nosample optimisation was interesting to me because the many independent trials show something about whether or not the optimiser gets stuck in local minima (which it does). With a single unified optimisation it's impossible to see "what you don't know". Since it hasn't improved in a few weeks, I've declared 6Xc2_nosample converged and put up a new Linac900Ext1Xc2_nosample with fewer parameters and potentially fewer local minima. I want to see if more people end up in the same place on this easier optimisation. On Linac900Ext6Xc2_nosample it looked like almost every user was in their own area! |
Stephen Brooks 2014-01-10 20:21:12 | [DPC] White Panther is somehow submitting results under version [v4.47v]. That I think is a shortened version of 4.47_dev. Maybe he downloaded the debug version and is still running that? |
Dave Peachey 2014-01-10 21:51:01 | Per your comment "it looked like almost every user was in their own area", and albeit it's very early days yet, I seem to be establishing a similar pattern again with four different instances running in four distinct streams with best results of: - Muon% ~ 0.280; Mpts ~ 2000 - Muon% ~ 0.690; Mpts ~ 1680 - Muon% ~ 1.250; Mpts ~ 500 - Muon% ~ -0.16; Mpts ~ 70 In each case, the Muon% is increasing but the maximum Mpts are staying approximately the same. From the majority of other users' results, they all seem to be going in approximately the same direction within a single stream - high Muon%; low/medium Mpts - rather like my third stream (above). It will be interesting to see whether things change significantly over the next few weeks .... |
[DPC]white_panther 2014-01-11 13:18:14 | [DPC] White Panther is somehow submitting results under version [v4.47v]. That I think is a shortened version of 4.47_dev. Maybe he downloaded the debug version and is still running that?fixed it now. |
Stephen Brooks 2014-01-12 23:21:56 | Thanks. Not a project-breaking event by any measure, in fact it's reassuring the newer codebase still works right in the real world. Recent change of compiler to MinGW shook out a few "silent bugs" that were working before just by virtue of how the old compiler LCC worked. BTW, it's weird that the "alpha" versions came out in full on the graph like this one of 8Xc2 but not the v4.47dev or _dev. What does the version tag in your original results say? |
[DPC]white_panther 2014-01-13 08:47:05 | [v4.47_dev] <Linac900Ext6Xc2_nosample> 56. [DPC] White Panther 48921 3`044410.7 (0.171%) 0.471535 4d 22 [v4.47_dev] <Linac900Ext8Xc2> 55. [DPC] White Panther 37026 19`471963.5 (0.204%) 4.265967 6d 0 |
Stephen Brooks 2014-01-17 15:44:41 | Thanks. These new optimisations are increasing pretty fast - 5Xc2 is nearly at 3% already. |