Re: help installing functions from repository (was: Re: Built-in parallelization in optimization functions)

[Olaf Till]

On Thu, Jun 19, 2014 at 10:54:22AM -0700, mikepatton wrote:
> Has any progress been made with parallelizing the optim package? I'm using
> the function samin for my problem, and if it could be parallelized, it would
> increase the speed tremendously. 

I'm not aware of attempts to parallelize samin.cc.

Some parallelization has been done for backends of new frontends of
the optim package. Since optim-1.3.1, this parallelization includes a
backend for simulated annealing, whose algortithm is, however,
different from that of samin.cc. Principally, a similar
parallelization could be possible for samin.cc.

Due to the nature of the simulated annealing algorithms I'm aware of,
the benefit of parallelization is limited and depends on the rate of
acceptance of steps. (The parallel results must be ordered, and only
those up to and including the first accepted can be used, the others
have to be discarded.) As long as many steps are accepted,
parallelization will only slow down things, even if the objective
function is computationally intensive.

Do you have a reference for a simulated annealing algorithm in which
parallelization can be more effective?

Olaf

-- 
public key id EAFE0591, e.g. on x-hkp://pool.sks-keyservers.net

signature.asc
Description: Digital signature