objective function is undefined at initial point. Fmin cannot continue

Hello,

I am trying to run a code and it always gave me the same error message:

"Error using sfminbx (line 27)

Objective function is undefined at initial point. fmunc cannot continue

I really need some help with my initial starting values or gmm objective function.

Any suggestions are welcomed!

Thanks!

global happycount invA ns x1 x2 s_jt IV theti thetj theta1 theta2 rho cdid cdindex nestid nestindex mktnestindex brand  


load ps2
load instruments


IV=[x1(:,2:end) instruments]; 
clear instruments inst

N=size(x1,1);

invA = inv([IV'*IV]/N); 

ns = 200;       
    
%starting values:
theta2w=    [0.09   0         0          0  0      0         ; 
             0.04   0      0.0228        0  0      0         ;
             0.06   0         0          0  0      0.0336      ];
          
             
[theti, thetj, theta2]=find(theta2w);

temp = cumsum(s_jt);
sum1 = temp(cdindex,:);
sum1(2:size(sum1,1),:) = diff(sum1);
outshr = 1.0 - sum1(cdid,:);

y = log(s_jt) - log(outshr); 




first=1;
for i=1:size(nestindex,1)
    last=nestindex(i);
    n = last - first + 1;
    s_jgt(first:last,:) = s_jt(first:last,1)./(ones(n,1)*(sum(s_jt(first:last,1))));
    first=last+1;
end


lnSjgt = log(s_jgt); 

rho=0.2;


delta_NL = y - rho*lnSjgt; .
mvalold = exp(delta_NL/(1-rho));
oldt2 = zeros(size(theta2));


save mvalold mvalold oldt2

options = optimset('GradObj','on','MaxFunEvals',900000);
clc

 
beta =[rho;theta2];

%Minimization
exitflag=0;
happycount = 0;
while exitflag==0
%[beta, fval,exitflag]=fminsearch(@gmmobj,beta,options) % Simplex search method  
[beta, fval,exitflag]=fminunc({@gmmobj,@gradobj},beta,options); % Newton method 
end
 

Answer 1

[Walter Roberson]

https://www.mathworks.com/help/optim/ug/fminunc.html#butpb7p-fun says

if you can compute the gradient of fun and the SpecifyObjectiveGradient option is set to true, as set by

options = optimoptions('fminunc','SpecifyObjectiveGradient',true)

then fun must return the gradient vector g(x) in the second output argument.

However, you syntax

{@gmmobj,@gradobj}

does not return the gradient vector in the second output argument: instead it tries to specify fun as a cell array of function handles.

You need

[beta, fval,exitflag]=fminunc(@(x) deal(gmmobj(x),gradobj(x)), beta, options); % Newton method 

[ebdilden]

Thanks much for your answer!

But it gives me an error when I put your command to my code? deal is an unexpected expression

[ebdilden]

also, I succeeded to make this code work with a small sample size; but when I put the all dataset it does not work. Also, my matlab is having a hard time to understand what is "deal"?

What's @x? since my x values are my betas defined over there.

Thanks for your help!

[Walter Roberson]

I copied and pasted from what I posted to a copy of your code, and it was fine. The only problem I saw is that your line

delta_NL = y - rho*lnSjgt; .

has an unexpected period at the end.

[ebdilden]

I think it's becuase of the "dot" that I forgot to delete in the end. That delta_NL is just a scalar vector.

I have matlab2016 and it does not work with the deal

[Walter Roberson]

We would need example input files to test.

[ebdilden]

Sure! My data is too big. I have selected a small sample and it works- what could be the reason that in a big data (100k observation) I am having an initial value problem?

[Matt J]

Have you tested the objective function by itself (separate from fmincon) on the initial input? It should be easy to trap any errors in the debugger and to observe any problem with the output.

[ebdilden]

my obj function is a function of two other function files- it is hard to check it out. any other recommendations?

[Walter Roberson]

Tracking is a bit easier if you parameterize the functions and get rid of the global.

http://www.mathworks.com/help/matlab/math/parameterizing-functions.html