|
|
fminunc: how reliable is the minimum?
Posted:
May 27, 2012 2:25 PM
|
|
Hello all,
I wrote a GMM (generalized methods of moments) program with the following objective function ( I show you the most important lines)
u=@(phi) -phi(1,:)*Z0(t,:)'+sum((Rc(t,:)-kron(Rf(t),ones(1,L)))'.*(phi(2:L+1,:)*Z0(t,:)'))+(Rw(t)-Rf(t))'.*(phi(L+2,:)*Z0(t,:)');
h=@(phi) (Rtot(t,:)-kron(Rf(t),ones(1,m))).*kron((1-u(phi))',ones(1,m));
EPSILON=@(phi) kron(Z0(t,:)',ones(m+1,1)).*kron(ones(l,1),[u(phi)' h(phi)]');
W=eye(l*(1+m));
f= @(phi) EPSILON(phi);
g= @(phi) (1/T)*sum(f(phi)')';
Q= @(phi) g(phi)'*W*g(phi);
----->[phi_1,fval]= fminunc(Q,START,options);
QUESTION:
The total number of observations (T) equals to 4300 and it takes a long time to execute this program. Even when the program finishes, it might not give right results as two executions produce differnet results. Nonetheless, the objective function seems quadratic so it should have only 1 minimum.
Is there any way I can make sure that I find a global minimum? Should I resort to genetic algorithm techniques?
Thanks!
|
|
