optimx with BFGS uses optim, so you actually incur some overhead unnecessarily. And BFGS really needs good gradients (as does Rvmmin and Rcgmin which are updated BFGS and CG, but all in R and with bounds or box constraints).
>From the Hessian, your function is (one of the many!) that have pretty bad >numerical properties. With all 0s, Newton is spinning in his grave. Probably the gradient is small also. So the optimizers decide they are at a minimum. As a first step, I'd suggest - checking that the function is computed correctly. That is, does your function give the correct value? - try a few other points "nearby". Are any "lower" than your first point? - Use numDeriv and get the gradient (and possibly Hessian) at each of these nearby points. These steps may reveal either that you have a bug in the function, or that it is pretty nasty numerically. In the latter case, you really need to try to find an equivalent function e.g., log(f) that can be minimized more easily. For information, I'm rather slowly working on a function test suite to do this. Also a lot of changes are going on in optimx to try to catch some of the various nasties. These appear first in the R-forge development versions. Use and comments welcome. If you DO find a lower point, then I'd give Nelder-Mead a try. Ravi Varadhan has a variant of this that may do a little better in dfoptim. You could also be a bit lazy and try optimx with the control "all.methods=TRUE". Not recommended for production use, but often helpful in seeing if any method can get some traction. Cheers, JN On 08/13/2011 06:00 AM, r-help-requ...@r-project.org wrote: > ------------------------------ Message: 47 Date: Sat, 13 Aug 2011 01:12:09 > -0700 (PDT) > From: Kathie <kathryn.lord2...@gmail.com> To: r-help@r-project.org Subject: > [R] > optimization problems Message-ID: <1313223129383-3741005.p...@n4.nabble.com> > Content-Type: > text/plain; charset=us-ascii Dear R users I am trying to use OPTIMX(OPTIM) > for nonlinear > optimization. There is no error in my code but the results are so weird (see > below). When > I ran via OPTIM, the results are that Initial values are that theta0 = 0.6 > 1.6 0.6 1.6 > 0.7. (In fact true vales are 0.5,1.0,0.8,1.2, 0.6.) > -------------------------------------------------------------------------------------------- >> > optim(par=theta0, fn=obj.fy, method="BFGS", control=list(trace=1, >> > maxit=10000), hessian=T) > initial value -0.027644 > final value -0.027644 > converged > $par > [1] 0.6 1.6 0.6 1.6 0.7 > > $value > [1] -0.02764405 > > $counts > function gradient > 1 1 > > $convergence > [1] 0 > > $message > NULL > > $hessian > [,1] [,2] [,3] [,4] [,5] > [1,] 0 0 0 0 0 > [2,] 0 0 0 0 0 > [3,] 0 0 0 0 0 > [4,] 0 0 0 0 0 > [5,] 0 0 0 0 0 > -------------------------------------------------------------------------------------------- > > When I ran via OPTIMX, the results are that > > -------------------------------------------------------------------------------------------- >> > optimx(par=theta0, fn=obj.fy, method="BFGS", control=list(maxit=10000), >> > hessian=T) > par fvalues method fns grs itns > conv KKT1 KKT2 xtimes > 1 0.6, 1.6, 0.6, 1.6, 0.7 -0.02764405 BFGS 1 1 NULL 0 TRUE > NA 8.71 >> > > -------------------------------------------------------------------------------------------- > > Whenever I used different initial values, the initial ones are the answer of > OPTIMX(OPTIM). > > Would you plz explain why it happened? or any suggestion will be greatly > appreciated. > > Regards, > > Kathryn Lord > > -- > View this message in context: > http://r.789695.n4.nabble.com/optimization-problems-tp3741005p3741005.html > Sent from the R help mailing list archive at Nabble.com. > > ______________________________________________ R-help@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.
