MilkyWay@home Project News: Research Update

[LAURENU2]

Things have been a little quiet for awhile so here's an update as to what we've been working on: 1. Nate has committed the convolution code, and we've finished debugging it. Hopefully today there'll be some new binaries as soon as I can get Joe to take a look at it (unfortunately he's the only one that can sign them at the moment). The convolution should take about 4-8x more time per work unit (which should make the people who want longer work units happy), because instead of calculating the fitness of the astronomy model against fixed star points, it calculates the fitness against probabilistic distributions of star points (ie, instead of a star being a fixed point, it has a probability distribution in a range of space). Nate is going to make a post about what exactly is going on with this in the next couple days. 2. The particle swarm searches seem to be running, however they're not performing nearly as well as I expected, so i'm looking into ways to optimize this. 3. I implemented a hybrid simplx/genetic search optimization and have been testing it out. The new genetic search work units you've been seeing are doing this. Where our previous genetic search reproduced parameters based on the double shot approach (see our escience paper), these generate child parameters using the simplex method. This is interesting because instead of generating child parameters based on two parents, the simplex method can use N parents. Right now i'm testing 4 (the next batch will test 5) and i'll keep increasing to see what effect N has on the convergence rate. 4. I'm working on implementing a simulated annealing optimization, so if you start seeing work units starting with 'sa' it will be for that. So that's what we've been up to in the last few days, thanks to everyone for continuing to crunch our numbers :) --Travis

More... (http://milkyway.cs.rpi.edu/milkyway/all_news.php#35)