Steve
Posts:
93
Registered:
8/4/08


Re: Problem with "lsqnonlin" function in matlab
Posted:
Sep 6, 2011 12:34 PM


"fsalehi " <fsalehi@gmail.com> wrote in message <j3nb0t$2ol$1@newscl01ah.mathworks.com>... > spasmous <spasmous@gmail.com> wrote in message <d247a3747fc44772b19bd800332e51cd@t30g2000prm.googlegroups.com>... > > On Aug 30, 12:34 am, "fsalehi " <fsal...@gmail.com> wrote: > > > I have a problem with 'lsqnonlin' function in matlab. I used it to optimizing this function: > > > > > > objfun = @(z) y  ( z(5).*exp(z(1).*(tz(7)).^2).*cos(2*pi.*z(2).*(tz(7))+z(3)) + z(6).*exp(z(1).*(tz(8)).^2).*cos(2*pi.*z(2).*(tz(8))+z(4)) ); > > > > > > % y is my signal and I optimize a guassian pulse function for Model Base Estimation > > > > > > options = optimset(options, 'MaxFunEvals', 10000, 'TolX', 1e3, 'TolFun', 1e2, 'MaxIter', 800, 'LargeScale', 'on', 'LevenbergMarquardt', 'on', 'DiffMaxChange', 1e7, 'DiffMinChange', 1e10, 'Display', 'off', 'Diagnostics', 'off'); > > > > > > > i would try Display 'iter' to see what is happening. TolFun seems kind > > of large and DiffMaxChange kind of small. Aslo I'm not sure > > LargeScale, LevMarq and bound constraints are all compatible either. > > I'm not sure too. but I dont know what setting I should use?
Regarding the LargeScale and LevenbergMarquardt settings, lsqnonlin will run the LargeScale (a.k.a. trustregionreflective) algorithm since the parameters are constrained by bounds. This is not an issue.
I agree about the DiffMaxChange being too small. When your parameter (alpha) is on the scale of 1e13, then a difference of 1e7 will be far too small to get an accurate finitedifference derivative estimate. What might be happening is that the derivative estimate for the first parameter may appear flat:
>> (1e13+1e7)  1e13 ans =
0
Try scaling the alpha parameter to be the same scale as the other parameters. Unfortunately, DiffMaxChange and DiffMinChange are onesize fits all options so increasing DiffMaxChange to a larger value may negatively affect your other parameters.
Regards, +Steve

