I am maximizin a log-likelihood function in an empirical econometric research to estimate a mixed logit model. For this purpose I use GlobalSearch with which I try to find the estimate of a theta vector of parameters (coefficients in the nonlinear regression that maximizes the log-likelihood). For quite some time I used theta_ig as an inital guess vector without being aware of that I am able to specify lower and upper bounds to the estimate of the theta vector. My idea was to specify these bounds in case that this helps to cut the time of minimization. When I use the bounds, however, GlobalSearch do somethings I could not quite understand. When I run GlobalSearch without the bounds, it returns an estimated theta with the note "fmincon stopped because the size of the current step is less than the value of the step size tolerance and constraints are satisfied to within the value of the constraint tolerance." So this is fine. When I specify bounds, as in the code below, however, GlobalSearch (i) takes more time to return a minimizing theta vector, (ii) it does not return the same minimizing theta vector as when I do not use the bounds, and (iii) it says "all 38 local solver runs converged with a positive local solver exit flag" which I do not get when I do not use the bounds. Despite reading how GlobalSearch is searching for the minimum quite a bit, I cannot seem to figure out why these three things happen.
options = optimset('GradObj','off','MaxFunEvals',10000,'Display','on','DerivativeCheck','off','Algorithm','sqp','TolFun',1e-9,'TolX',1e-9,'Display','iter');
problem = createOptimProblem('fmincon','objective',obj,'x0',theta_ig,'lb',theta_lb,'ub',theta_ub,'options',options);
gs = GlobalSearch('Display','iter');
[theta_hat,fval,exitflag,output,solutions] = run(gs,problem);
