I'm afraid I don't have any ideas. I suppose that would be a question for the 
Matlab forums.

-- 
Ray Zimmerman
Senior Research Associate
211 Warren Hall, Cornell University, Ithaca, NY 14853
phone: (607) 255-9645



On Mar 31, 2011, at 9:00 PM, m d wrote:

> Thanks a lot for your response and also the reference.
> 
>  I read the algorithm. It hase been said that fmincon uses a quasi-Newton 
> approximation to the Hessian by the BFSG formula, as you mentioned. However, 
> the hessian is the second derivative of the Lagrangian function w.r.t  X. So 
> it should be close to the hessian provided by Lxx in mips, since Lxx is the 
> accurate second derivative of Lagrangian function. 
> Can you explaine me why there is such a large difference as follow:
> 
> Hessian provided by fmincon =
> 
>     2.2313    8.7213
>     8.7213   60.3726
> 
> 
> Hessian (Lxx) provided by mips  =
> 
>    -0.3666         0
>        0             -0.0046
> 
> Please note that the optimal solution given by both solvers, mips & fmincon, 
> are equal, just different in their supplied hessian.
> 
> Thanks in advance
> 
> 
> --- On Wed, 3/30/11, Ray Zimmerman <[email protected]> wrote:
> 
> From: Ray Zimmerman <[email protected]>
> Subject: Re: mips & fmincon
> To: "MATPOWER discussion forum" <[email protected]>
> Date: Wednesday, March 30, 2011, 4:17 PM
> 
> Based on the fmincon documentation, fmincon has a number of ways of handling 
> the Hessian, depending on the algorithm. In many cases it uses a quasi-Newton 
> approximation to the Hessian. I suspect that's what you're seeing.
> 
> -- 
> Ray Zimmerman
> Senior Research Associate
> 211 Warren Hall, Cornell University, Ithaca, NY 14853
> phone: (607) 255-9645
> 
> 
> 
> On Mar 30, 2011, at 3:46 PM, m d wrote:
> 
>> Hello all,
>> 
>> I've  faced with an interesting question and appreciate if anyone can answer!
>> 
>> When I use "mips" as an optimization solver to solve one of those simple 
>> optimization problems with constraint in t-mips.m, I get a hessian of the 
>> lagrangian by calling " Lxx" . The hessian provided in Lxx is closely 
>> matched with that I compute the problem by hand. However, when I use 
>> "fmincon" to solve the same problem as following:
>> 
>> [x,fval,exitflag,output,lambda,grad,hessian] = fmincon(...)
>> 
>> and compare its hessian with Lxx, I observe a huge diffence, though the 
>> optimal answers for x and objective function are the same.
>> 
>> Can anyone answer why? What is the algorithm used in fmincon to compute the 
>> hessian?
>> 
>> Thanks in advance for your reply,
>> Maryam
> 

Reply via email to