Problem parallelizing a nested optimization problem

Dear all,

I am having trouble parallelizing the following nested optimization problem. In the upper level nest I maximize a function, in search for the optimal parameter vector (c_par). Each iteration of this upper level optimization level involves a series of 132 (potentially more) independent optimization problems. Since running the program serially takes a very long time, I've been trying to parallelize the lower level optimization problems with PARFOR loops.

I keep getting the following error: A parfor body cannot refer to a nested function (foc_par1).

Here's what I'm doing, and maybe anyone has suggestions that can help me solving this in parallel:

Upper level optimization called from upperlevel.m

        [cpar_ols, fval, exitflag] = fminunc('conf_game_par_new', cpar_start, options1);

The m-file conf_game_par_new.m then does several things (given a vector of parameters c_par):

start=0;

parfor s=1:132
%%A block of data transformations and setup of the s-optimization problem

        e_init   = par;
        cost_par = Alpha_p(1:5);         
        Alpha_init = 0.2.*ones(n,1)';
        [Alpha] = fminunc(@foc_par1, Alpha_init, options);
end

Here's where I get my problem, since the function foc_par1 used in the lower level optimization problem for each iteration s, and is defined at the same in the same m-file.

I'm not sure what to do here.

Thanks in advance.

Daniel

Answer 1

[Jiro Doke]

That happens to be one of the restrictions for PARFOR. You should either make function foc_par1 a subfunction or a separate function MATLAB file.

Answer 2

[Daniel]

Hi Jiro, thanks for your reply. It appears to be working now. I believe that the problem was on how to handle each s-specific data, which I ended up passing as extra parameters.

Thanks, Daniel