[R] sporadic errors with nlrq() / optim()

[Peter Dalgaard]

"Rogers, James A [PGRD Groton]" <James_A_Rogers at groton.pfizer.com> writes:

> Dear List, 
> 
> Apologies if this is a known problem ... I wasn't able to find it on the bug
> list, but it is a problem that does not seem to occur with a MAC build of R
> 2.0, so perhaps this problem has already been addressed for the future.  
> 
> I am getting *sporadic* errors when refitting the same model to the same
> data set, using nlrq() in the nlrq package. The algorithm is not stochastic,
> so I would expect to get errors either every time, or never.
> 
> ###
> 
> library(stats) # or library(nls) if using R < 1.9.0
> library(nlrq)
> 
> test <- data.frame(x = c(7.60090245954208, 6.90775527898214,
> 6.21460809842219, 5.52146091786225, 4.60517018598809, 3.91202300542815,
> 3.2188758248682 , 2.52572864430826, 1.83258146374831, 7.60090245954208,
> 6.90775527898214, 6.21460809842219, 5.52146091786225, 4.60517018598809,
> 3.91202300542815, 3.2188758248682 , 2.52572864430826, 1.83258146374831,
> 6.21460809842219, 6.21460809842219),
>                    y = c( 11.0161506644269, 9.84267541313937,
> 8.66146668057266, 7.48099216286952, 6.50578406012823, 6.24027584517077,
> 5.63121178182137, 5.71702770140622, 5.64190707093811, 10.8983676287705,
> 9.91857318995417, 8.74608021735751, 7.58120982619635, 6.361302477573 ,
> 5.91889385427315, 5.63835466933375 , 5.80211837537706, 5.64897423816121,
> 8.6195692580331 , 8.70367275835886)
>                    )
> 
> i <- 1
> while(i < 500) {    # I usually hit an error within 50 iterations
>   cat(i, "\n")
>   nlrq(y ~ SSfpl(x, A, B, xmid, scal), data = test)
>   i <- i + 1
> }
> 
> ###
> 
> Errors occur with version 1.8.1 and 1.9.0 on both Windows (two different
> machines) and UNIX, but not on version 2.0 on a MAC (these are the only R
> version - OS permutations I was able to get reports on easily). 
> 
> Anyone understand what is happening here?

There's no "version 2.0", only "2.0.0 Under development (unstable)". 
If you look at the NEWS file for that version (or for 1.9.0-patched)
under 1.9.1 changes, you'll find

   o   The L-BFGS-B option of optim() apparently needs part of its
        workspace zeroed.  (PR#6720)

which gave problems of the sort you experiences.

-- 
   O__  ---- Peter Dalgaard             Blegdamsvej 3  
  c/ /'_ --- Dept. of Biostatistics     2200 Cph. N   
 (*) \(*) -- University of Copenhagen   Denmark      Ph: (+45) 35327918
~~~~~~~~~~ - (p.dalgaard at biostat.ku.dk)             FAX: (+45) 35327907