Use of an anonymous functions with additional parameters within a spmd-block

Hi community,

this is the first time I couldn't find a solution for my problem by searching the forum, so i decided to post it.

I have a bigger function which uses fminsearch with an anonymous function + additional parameters in a spmd-block. I don't know how to pass these additional Data to my user-fct.

For understanding my problem I generated a minimal-expample. (I know parallel computing makes no sense for this optimization, but my problem is much bigger and takes a long time to calculate):

function []  = start()

       for i=1:10

           p=i*10;
           min = fminsearch(@(x)target(x,p),[10])

       end

    %function to be minimized
    function [val] = target(x,p)
        val = (x-p)^2;
    end

end

Now, I want to calculate the for-loop in several threads. So I decided to use a spmd-block (also tried parloop,...). I get something like this:

function []  = start()

    global p;
    x = [];
    p = [];
    fh_target = @(x)target(x);

    matlabpool 5;
    spmd
       for i=labindex:numlabs:100

           p=i*10;
           min = fminsearch(fh_target,[10])

       end
    end
    matlabpool close;

    %function to be minimized
    function [val] = target(x)
        val = (x-p)^2;
    end

end

But i dont know how to pass the parameter p to the function target via a function-handle that is used by fminsearch. As you can see i tried it with global variables, but this doesn't work.

The parameter is not know before the for-loop or the spmd-block. I'm sorry for my english, but i hope aou understand my problem and somebody can help me.

Best Regards

Answer 1

[Edric Ellis]

You need to work around the fact that SPMD blocks cannot call internal functions or define anonymous functions. I would do it like so:

    function mn = start()

    spmd
        % Use a codistributed array to store results
        mn = zeros(1,100,codistributor());

        % Loop over the elements of the codistributed array in parallel
        for i = drange(1:100)
            p=i*10;
            % SPMD blocks cannot call nested functions, nor declare anonymous functions, so
            % call helper function to do optimisation.
            mn(i) = doSolve(p,10);
        end
    end
    % Gather the distributed 'mn' back to the CPU.
    mn = gather(mn);
    end

    % Helper function to call fminsearch
    function mn = doSolve(p,v)
    mn = fminsearch(@(x)target(x,p),v);
    end

    %function to be minimized
    function [val] = target(x,p)
    val = (x-p)^2;
    end

Answer 2

[Andreas]

Thanks a lot, works perfect.