Parallel Optimization: Loop or Finite-Difference Gradient

According to the documentation on parallel optimization, one cannot use the 'UseParallel' option inside a 'parfor' loop.

I am optimizing a very messy likelihood function over 32 parameters w/ fmincon. Unfortunately I run the optimization around 1,000 times each at a different initial point.

Would it be more computationally efficient to parallelize the finite-difference gradient calculation, or parallelize the loop of 1,000 calls to fmincon.

Thanks!

PS I have 25 cores available to me.

[Walter Roberson]

Is what you are doing really a global optimization? If so then the tools such as MultiStart in Global Optimization Toolbox would seem appropriate. Or patternsearch.

Answer 1

[Alan Weiss]

It is very difficult to know which way would work better. Can you test a small set of representative points and see which way is faster?

Sorry for the lack of info, but parallel processing is hard to predict, as there are a lot of things that go into its performance, or lack thereof.

Alan Weiss

MATLAB mathematical toolbox documentation