[R] OPTIMX: non-finite finite-difference value [26] and scaling problem

[Bert Gunter]

... but unless there is an outright error in the code, the problem is
due to the specific data and starting values, which means they cannot
be easily reproduced.

More than likely, the optimizer has run into numerical problems. After
all, it is wandering around in 37 dimensional space and the max/min
parameter ratio (scaling issues) is  about 1000. So there could well
be severe overparameterization/inadequate data and poor parameter
scaling. Many applied practitioners seem unaware of the realities and
challenges of *non*linear optimization -- it is, after all, a highly
technical specialty -- and somehow expect that they can throw
multi-parameter nonlinear models at any old data they have. This is
the sort of rude awakening that might be expected under such
circumstances, for which the only solution may be a careful rethink of
the goals of the modeling effort, the complexity of the models, and
the limitations of the data available.

Note that these comments are of necessity largely speculative and
would benefit greatly by amplification and criticism by real experts,
of which I ain't one. So caveat emptor! (as usual on the internet).

Best,
Bert


Bert Gunter

"Data is not information. Information is not knowledge. And knowledge
is certainly not wisdom."
   -- Clifford Stoll


On Sat, Jul 18, 2015 at 12:07 AM, Jeff Newmiller
<jdnewmil at dcn.davis.ca.us> wrote:
> It is quite unlikely that anyone can help you without a reproducible
> example. [1]
>
> You should be aware that sending emails to this list in HTML format is
> likely to yield scrambled mails received. Please make your email client send
> email in plain text.
>
> [1]
> http://stackoverflow.com/questions/5963269/how-to-make-a-great-r-reproducible-example
>
> On 07/17/2015 10:55 AM, Olu Ola via R-help wrote:
>>
>> Hello,I am running a nonlinear GMM using the optimx wrapper. I am trying
>> to estimate 37 variables however and my code for the optimx is:
>> nlgmm = optimx(par=b0, fn=obj,method = "BFGS", itnmax=10000,
>> control=list(follow.on = TRUE,kkt=FALSE,starttests=TRUE,save.failures=TRUE,
>> trace=0))
>>
>> My staring values are from Ordinary least squares estimates (OLS) and they
>> are:
>> b0 <-
>> c(-2.00658,-0.04373,0.19079,0.34652,-0.36814,0.21284,-0.24369,0.64622,0.22927,0.29431,0.19547,0.80614,18.8398,0.5928,3.1375,0.4301,-0.4937,2.2016,31.5203,0.6171,1.0206,1.7830,-0.4421,11.0076,-0.03305,0.17087,0.44794,0.17488,0.10781,-0.50747,-0.04563,0.17030,0.41792,0.17526,0.99734,-17.2996,-41.9359)
>>
>> I got the following error:
>> Error in optim(par = par, fn = ufn, gr = ugr, lower = lower, upper =
>> upper,  :   non-finite finite-difference value [26]
>> I also got an error about scaling which is as follows:
>> Parameters or bounds appear to have differentscalings.This can cause poor
>> performance in optimization.  Itis important for derivative free methods
>> like BOBYQA, UOBYQA, NEWUOA.
>>
>>
>> any help will be greatly appreciated.
>> Best Regards
>>         [[alternative HTML version deleted]]
>>
>> ______________________________________________
>> R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see
>> 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.
>>
>
>
> --
> ---------------------------------------------------------------------------
> Jeff Newmiller                        The     .....       .....  Go Live...
> DCN:<jdnewmil at dcn.davis.ca.us>        Basics: ##.#.       ##.#.  Live Go...
>                                       Live:   OO#.. Dead: OO#..  Playing
> Research Engineer (Solar/Batteries            O.O#.       #.O#.  with
> /Software/Embedded Controllers)               .OO#.       .OO#.  rocks...1k
>
> ______________________________________________
> R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see
> 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.