waffleironhead 2007-06-07 20:07:30 | Man oh Man is this exciting. DecayRotB has been on the climb. In just 2 months it has doubled its efficiency. ~4.1 to ~8.2%. Lets hope the next 2 months are just as eventful. Congrats to everyone who has been crunching. |
RGtx 2007-06-07 21:00:12 | I think that you may have displaced the decimal point a little. Nevertheless, positive progress is being made in both projects. |
Stephen Brooks 2007-06-08 08:51:45 | Yep, pretty encouraging, especially as the thing that broke it out of its rut was a manual seeded design, but that only gave some tiny (5% relative) increase and now it's been optimised to double. After the conference this summer I may be taking that apart and seeing exactly what it was about the manual design that made it so promising, perhaps with a view to integrating that sort of thing into the optimiser at a later stage. |
Stephen Brooks 2007-06-08 10:32:26 | Interesting also: this is the lattice that I used a "fixed random seed" on. So the optimisation gradient should be more obvious to the program. |
Stephen Brooks 2007-06-08 10:39:56 | Right, in that case I'm going to switch determinism on in PhaseRotD, giving a new lattice "PhaseRotDD". Please would someone do a manual seed on DD of the best result in the samplefile from D? Because basically it should start off from the same yield point, with a small change due to the random seed. |
Stephen Brooks 2007-06-08 10:49:30 | Actually, I think I might do it. Noticed a few solenoids I wanted wider as well, should give a good starting point... |
RGtx 2007-06-08 11:06:53 | One point to bear in mind, is that with a fixed 'random' seed, the program will optimize in the direction where the mean yield of five runs is greater than that of the first run alone. |
RGtx 2007-06-08 11:21:44 | Sorry, for "greater than..." read "lower than that of the first run alone". |
Stephen Brooks 2007-06-08 11:39:29 | The rule for re-checking is that anything that exceeds the current highest result with #runs=5 will also be re-checked. So in the very highest results, it will always run the first five (also fixed) seeds. |