 |
| [AMD Users] Michal Hajicek 2004-05-26 11:18:47 | Hi Stephen, thank you, it's good that everyone can now set how often different trial-types will be used. But, could you explain the new ones-how they work, please.(Extrapolate, MuSpherical, MuOne, Extreme, LocalGrad, TopoSmooth) And, what's 'particles per extra thread'? Michal |
| Stephen Brooks 2004-05-26 11:27:29 | Extrapolate is like interpolate except it takes the line between two good results and selects a point on the _exterior_ part of that line, so both of the original results lie on the same side of it. MuSpherical makes a random jump within a sphere centred on a good design (normal Mutate is more of a rectangular star pattern where only a few parameters at once will be varied). MuOne only mutates one parameter at a time. Despits its simplicity, this one did quite well in the tests I did (though not as well as Extrapolate and LocalGrad). Extreme makes some of the parameters closer to their end-values. This is designed for optimisations where a lot of parameters may be at their maximum or minimum. LocalGrad is more clever: it looks at points within a certain distance of a particular good result and estimates the average slope of muon% in that region, then steps 'uphill' by a certain distance. TopoSmooth actually takes account of the parameter names: it treats two with consecutive numbers (e.g. s12l, s13l) as neighbours and then 'smooths' the parameters, so each is closer to its neighbour' values. This is designed to be good in optimisations where the parameters vary smoothly along the machine. Particles per extra thread is just the number of particles required for Muon1 to justify starting an additional thread. Below a certain number of particles, having multiple threads is more costly due to the O/S overhead of creating them than just doing the calculation directly in one thread. |
| AySz88 2004-05-26 19:59:31 | Why would it be faster with extra threads? Does it matter for single-processor machines? |
| Herb[Romulus2] 2004-05-27 05:39:18 | Is it possible to set Random=0 ?? |
| Stephen Brooks 2004-05-28 01:12:45 | quote:Only hyper-threaded ones. If you've set threads=1, then the particles-per-extra-thread setting does nothing. Obviously.  quote:Random trials are added according to a probability that decreases with number of results in results.dat. This tends to zero as the number of results becomes large. However, if you want more random results than that, you can set Random=1; or similar in that line. |
| Herb[Romulus2] 2004-05-28 10:55:52 | No, I just wanted to exclude randoms completely for some testing purposes, as they are filling up the results.dat when I tried to optimize between only 2 given fine breed results through all the other strategys. Randoms in such a case are just watering the results. No big deal, I can replace the results.dat every hour by my own sample.dat then |
| kitsura 2004-05-28 17:20:08 | So is 0 considered a valid range for the trial types or is it only from 1-3? How about 4 and above? |
| Herb[Romulus2] 2004-05-28 23:00:13 | I think it's just a relationship number, but it didn't work with Random=0 for me, all others with 0 just worked fine. A setting up to maximal 9 for a method should be safe. It will tell the client to do this method 9 times more often than one with a setting of 1. I'll check it, but it will last at least half a day and if I run into reruns, even longer |
| kitsura 2004-05-29 04:42:48 | Same thing here I haven't gotten Random=0 to work. |
| Pollock[Romulus2] 2004-05-29 07:26:14 | It seems that the Random runs are just a curse we will need to live with.  The 'Extreme' runs are nice, though. Those runs usually 'max out' parameters that I was planning to do anyway. It saves me the trouble of doing it manually. |
| Herb[Romulus2] 2004-05-29 11:03:43 | 60630000666606006666 that was the gen-pattern with one rerun results, so it behaved like expected. I'll try tomorrow a different pattern with random=1 |
| Stephen Brooks 2004-10-20 11:31:58 | It's just divided down - I think it actually accepts values up to 999, while you can't at present turn off "random" completely. |
