Re: Behaviour of FMINCON - question.

Vijit,

Thank you for your reply. Any comments appreciated.

""As I understand it, your solution is getting trapped in local
minima. fmincon uses a simplex method for optimization which does not
have the ability to recover from a local minima.""

I am 99.99% sure that this is not the case. (1) I believe that there
is only one minimum for this particular objective function. (2) If it
was a local minimum, then surely fmincon would satisfactorily
converge (noted by exitflag). Instead, even with a very large max
number of function evaluations, it doesn't converge. And (3) I have
actually checked to see if the point FMINCON "stops at" is infact a
minimum. It isn't. If you substitute the numerical values for 2
parameters at a time into the objective function, and then plot the
objective function versus the third parameter (2-D plot), you see
that the minimum of this plot does not occur at the point which
FMINCON says it should.

""Whenever the search criteria is not met, repeat your search with
different initial guesses (kind of a brute force method).""

I have done this. Sometimes FMINCON does converge on more than one
occasion to one point - which is a minimum. Other times, FMINCON
"gets stuck", as my original post explained, but does not "get stuck"
at a point which corresponds to a minimum. The point it gets stuck
at is not always the same point. Again, these different points have
been checked, and do not sorrespond to local minimia.

"John has written a scrip for "bound constrained optimization using
fminsearch".

I shall try this - thank you. However, it still doesn't solve my
current problem/query regarding FMINCON.

Again, Vijit, thank you for taking the time to offer advice.

Regards,
Barry

vijitnair@xxxxxxxxx wrote:


As I understand it, your solution is getting trapped in local
minima.
fmincon uses a simplex method for optimization which does not have
the
ability to recover from a local minima. A couple of solutions to
this
may be
1. Whenever the search criteria is not met, repeat your search with
different initial guesses (kind of a brute force method).
2. Try another constrained optimization search. John has written a
scrip for "bound constrained optimization using fminsearch". This
script (fminsearchbnd) is available on
MCFE>http://www.mathworks.com/matlabcentral/fileexchange/loadFile.do?objectId=8277&objectType=file <http://www.mathworks.com/matlabcentral/fileexchange/loadFile.do?objectId=8277&objectType=file>
3. Use other optimization methods like simulated annealing that
have
hill climbing abilities so as to recover from local minima.

HTH
Vijit


.

Relevant Pages

  • Re: Scaling in fmincon()
    ... Yes, yes, I did change the tolerance values accordingly. ... optimization variables are of similar magnitude, ... imply that Matlab fmincon() does not do these. ... I'll form two objectives, ...
    (comp.soft-sys.matlab)
  • Re: Optimal Power Flow with fmincon
    ... I don't know anything about OPF, but I have used fmincon a fair amount. ... Depending on what my objective function looked like, I sometimes had to change some of the default optimization parameters. ... The other thought that I have is that if the different variables of the objective function are scaled very differently you might have problems. ...
    (comp.soft-sys.matlab)
  • Re: Need help with fmincon and fminunc
    ... fmincon,and when I run the optimization program,this ... The default in fmincon is to use the largescale ... 1.I don't have any constraint function.However,I have the ... Fminunc has no ability to use constraints. ...
    (comp.soft-sys.matlab)
  • optimization
    ... fminsearch, fmincon and fminimax. ... optimization with 1 parameter, so I have tried also with ... in each iteration, and a minimum is obtained at X1=136,5956. ... I have tried with fminsearch. ...
    (comp.soft-sys.matlab)
  • Re: minimizing a function in matlab in a range
    ... how do i go about it, i tried fmincon but looks like it can solve only at a point ... optimization tool like fmincon will be adequate to ... internal optimization with fminbnd is only accurate ... to your supplied tolerance. ...
    (comp.soft-sys.matlab)