randomwalk with parallel optimization

[wmliu]

Hello,
I'm trying to do parallel optimizations on a quad core desktop and trying to figure out which method is the best for job.
From the manual, it says that the method for parallel optimization may be one of "genetic", "hybridsimplex" or "swarm". I don't have much computation resource, so "swarm" is out of consideration. My start point is no where closed to the solution, so "hybridsimplex" is out. And "genetic" doesn't seems to be working for me either.
I tried to use the default method and it does work but it will give me a local optimium some time depends on my start point. If I put down "randomwalk" for the method, will Pelegant do the optimization as "randomwalk" in serial mode except tracking in parallel?

Thanks,

Wanming

[michael_borland]

Wanming,

If you are nowhere near a solution, randomsample followed by randomwalk is recommended. If you use &optimization_setup, elegant will track in parallel, so you'll get the benefit of your four cores.

--Michael

[ywang25]

In the parallel optimization methods, both swarm and genetic are supposed to be global optimization methods. The hybridsimplex should give a result at least better than the default (serial simplex) method with the same configuration, where an independent simplex method will be conducted on each of the cores. Even you have a limited number of the cores, you can still try the swarm method with the population size (e.g. 100) larger than the number of cores. The example given in the paper with a large number of cores for the swarm method is designed to be finished in a very short time period. If the quality is your high priority, you can either try the serial method suggested by Michael, or the parallel swarm method with a limited number of cores.

Yusong

[wmliu]

Michael and Yusong

Thanks for the help.

Wanming