parfor with fminunc where fminunc returns a vector

Hi,

I am running a parfor loop inside which I use fminunc and I get the following error:

Starting matlabpool using the 'local' profile ... connected to 12 labs.
Error using parallel_function (line 589)

User supplied objective function must return a scalar
value.

Error stack:
fminunc.m at 324

Error in untitled8_noprices (line 72)
    parfor j=1:J    % loop for firms (column of matrices)

Is there some way to fix if I screwed something or to go around any limitations that I might be running into?

I attach my code under so that you can look at it/make suggestions

------------------------------------------------

while dif>tol && its<maxits 
    parfor j=1:J   
        for i=1:N   
            for ii=1:N
                k0=kmat(i,j);
                m0=mmat(ii,j);
                km0=[kmat(i,j),mmat(ii,j)];
                km1 = fminunc(@(km) firmprob_2cap_noprices(km,j,vold,k0,m0,Agrid),km0);%,options);
                vnew(i,ii,j) = -firmprob_2cap_noprices(km1,j,vold,k0,m0,Agrid);
                kopt(i,ii,j) = km1(1);
                mopt(i,ii,j) = km1(2);
                Ymat(i,ii,j)=Agrid(j)*(k0^alpha)*(m0^alpha)-km1(1)-km1(2);
            end         
        end
    end
    if its>=maxits;
        fprintf('Ran out of iterations \n');
    end
    for jj=1:J
        diff(1,jj)=norm(vnew(:,:,jj)-vold(:,:,jj));
    end
    dif=max(diff)
    vold = vnew;
    its = its+1;
end

[J]

sorry, i should mention that km1 is supposed to be an array with two elements... code works fine when i use a for loop as opposed to a parfor loop

[Walter Roberson]

We will need to see firmprob_2cap_noprices

[J]

function val=firmprob_2cap_noprices(km,j,vold,k0,m0,Agrid)

global beta delta alpha eta 

% do the interpolation
X=[km(1),km(2)];
gg=interpne(vold(:,:,j),X);

profit = (Agrid(j).*(km(1).^alpha).*(km(2).^(alpha)))-(km(1)-k0*(1-delta))-(km(2)-m0*(1-delta));

if profit<=0 
    val = -9999999 - 999*abs(profit);
else
    val= profit+ beta.*gg;
end

val = -val; % make it negative since we're maximizing and code is to minimize.

Answer 1

[Walter Roberson]

Is "interpne" from the Factors on Demand FEX contribution?

What are size(alpha), size(beta), size(delta) ?

Please put in a conditional breakpoint on the last line of the function, to stop if length(val) ~= 1

My speculation is that beta might be a vector rather than a scalar.

Note: you are safer passing alpha, beta, delta, eta into the function as parameters rather than as globals. Using global with parfor is unsafe if the functions can somehow write to the globals.

And of course there would be problems if the globals ended up initialized to [] because they had not been passed into the workspace of the worker.

[Walter Roberson]

See http://www.mathworks.com/matlabcentral/answers/33344-does-parfor-support-global-variables for explicit statement that global variable values are not copied to the workers.

Answer 2

[J]

Hi,

Even though alpha, beta and delta are scalar and not vectors (0.2,0.9 and 0.1 respectively), when i passed them through the function rather than as globals, the parfor worked!!!!!!!!!!

So greatful for your help, I really appreciate it!

I have an added question, my fminunc prints 'Local Minimum found'. Is there some way to stop that to quicken the code?