fmincon not using user-supplied gradient to optimize

Hi Everyone,

I am facing a problem using fmincon function. I am trying to use my user-supplied gradient instead of the one gained by finite difference method. But, the thing is fmincon is not using it even if I have supplied it as the second output in my objective function. And a portion of my code is below:

 %
options = optimoptions('fmincon','Algorithm','sqp','UseParallel', true,'Display','iter','GradObj','on','TolX',1e-4,'TolFun',1e-4,'DerivativeCheck', 'on');
[r,fval,exitflag,output] = fmincon(@Objfunc,lb+0.01,[],[],Aeq,beq,lb,ub,[],options)
function [ F, gradF ] = Objfunc(r)
% Some complex function omitted here
end

I have tried several method to make it work:

1. When I turn on 'DerivativeCheck' option, the result showed that the difference is larger than 1e-6. Therfore, it won't pass the check.

2. So, I looked into my user-supplied gradient and sort it from largest to smallest and compare to the optimization result using finite differnce. I am pretty sure now that my gradient is correct (Though the scale is very small which is possibly the reason why it couldn't pass the DerivativeCheck)

3. Therefore, I turned to edit the validateFirstDerivatives.m file to enlarge my tolerance.This time the DerivativeCheck passed and therefore fmincon used my user-supplied gradient and the result perfectly matched what I got without supplying it.

(Suggested by previous post: http://www.mathworks.com/matlabcentral/newsreader/view_thread/339795)

So, the question is: Can I avoid DerivativeCheck to enforce fmincon using my user-supplied gradient? Because DerivativeCheck need a lot of time to pass and I have a need to speed up my optimization process. Since the objective function takes time to compute, using the user-supplied gradient is necessary for me.

If someone could kindly tell me the way to solve this issue? Thank you for your reply.

[CP Chen]

Okay, I guess I just figured out what going on. I just set the DerivativeCheck option to off then everything is fine now. I guess there is some DerivativeCheck if I didn't set it explicitly. And therefore, if the user-supplied gradient is slightly off with the result by finite difference, fmincon will use finite difference anyway.