Failure of initial objective function

I am getting the dreaded "Failure in initial objective function evaluation. FMINCON cannot continue." Notification.

I am trying to use fmincon, but also using a simple loop within the function to see how parameters change over time.

Here is my code:

```

%Define Function- fmincon%

opts = optimset ('Display','iter','TolX',2e-9,'TolFun',2e-9,'MaxIter',100000,'MaxFunEvals',10000000,'FunValCheck','on');
[param,fval,exitflag,output,lambda,grad,hessian] = fmincon(@(param) Posterior(data,idx,param),param0,A,b,Aeq,beq,lb,ub,nonlcon,opts); 
err  = sqrt(diag(inv(hessian)));



%Run posterior function that is solved for Ep1.5%

k = 1;  %number of predictors


function LL = Posterior(data,idx,param)
%n = 13;  %number of obs: simple case
n = max(data.AgeMinus1); %Will update to go by ID

% define components of the function:
%EpPrior0  = param(idx.EpPrior);   

Tau0       =param(idx.Tau);       %Prior Variance/precision

Sigma_sq0  = param(idx.Sigma_sq); %Variance of 

ShockInfo0 =param(idx.ShockInfo); % n  = how informative the observation was

% Define vector to hold the shocks:
Shock = zeros(n, 1);
% Define a vector to hold the precision of the prior:
Tau      = zeros(n+1,1);
Tau(1,:) = Tau0(1,:); % Variance of prior going into period 1


for i=1:n % Start the sampling
    Shock(i,:) = data.Ep2(i,:) + (data.Ep2(i,:)- data.Ep1(i,:)).*(Sigma_sq0(i,:)./(ShockInfo0(i,:).*Tau(i,:))); %Ep 1.5
    
    Tau(i+1,:) = ((Sigma_sq0(i,:)./ShockInfo0(i,:)).*Tau(i,:))./(Tau(i,:) + (Sigma_sq0(i,:)./ShockInfo0(i,:)));
end
LL = -sum(log(normpdf(Shock, 0, Sigma_sq0)));

end

```

How can I fix this error?

Thank you in advance!

[Chris]

Try calling

Posterior(data,idx,param0)

Do you get a valid answer?

[Joshua Scott]

It doesn't recognize the function, which seems weird to me.

[Matt J]

Try calling it within your script, before the optimization is launched.

[Joshua Scott]

I get "Unrecognized function or variable 'param'."

Which is interesting, since I have ran ML and fmincon while defining 'param0' without 'param'.

Should I define param using the idx structure?

[Torsten]

You should define "param0" in the same way as you extract the fitting parameters from "param" in "Posterior".

[Joshua Scott]

Thank you! I did have two parameters swapped. Small progress!

Now I get the same error, but can call on the function at param0 and its not a number (NaN).

Answer 1

[Torsten]

Then set some outputs in the function "Posterior" to see where the NaN operation occurs.

[Joshua Scott]

There seemed to be an updating issue with the nested function. I changed param0 to vectors of ones to avoid issues. It runs! Now on to a new set of problems!

Thank you so much!