Dear useRs,
I have a complicated function to be optimized with optim(), and whose
parameters are passed to another function within its evaluation. This
function allows for the parameters to enter as arguments to various
probability distribution functions.
However, I am violating some scoping convention, as somewhere within the
hierarchy of calls a variable is not visible. I will give a genericized
example here.
> myFxn <- function(parms, Y, phi, <other args>) {<body of function>}
### I want to optimize this over its first argument
> optim(par=<numeric(>2)>, fn=myFxn, ### end of named args, next are
all in "..."
Y=<data>, <other args>,
phi=function(r) pnorm(r, mean=parms[2], sd=parms[1])
)
Error in pnorm(r, mean = parms[2], sd = parms[1]) :
object 'parms' not found
> debugger() ## options(error=expression(dump.frames())) in
.RProfile
Message: Error in pnorm(r, mean = parms[2], sd = parms[1]) :
object 'parms' not found
Available environments had calls:
1: optim(par = <numeric(>2)>, fn = myFxn, Y=<data>, <other args>, ..
2: function (par)
3: fn(par, ...)
4: ifelse(<logical vector>, phi(Y), <other stuff within myFxn>
5: phi(Y)
6: pnorm(r, mean = parms[2], sd = parms[1])
Now, when using the debugger in environments 1 and 2, I can see the
value of 'par' correctly. I cannot access environment 4 as it just
returns the original error message. Trying to access environment 3
gives the (to me) cryptic 'Error in get(.obj, envir =
dump[[.selection]]) : argument "..." is missing, with no default' and
returns to the top level without debugging.
I will try to explain to the best of my ability what I think is
happening here. Environments 2 and 3 are from the first lines of
optim(), where it is building an internal function to evaluate the
candidate parameter values. When accessing environment 3, it seems like
when it fills out the "..." argument of fn(), it is passing
phi=function(r) pnorm(r, mean=parms[2], sd=parms[1]) but upon trying to
evaluate the variable 'parms', it cannot see it in the search path.
When actually running the original call, 'parms' is apparently not
evaluated yet, but is once the pnorm call is hit. It appears the
'parms' variable is being evaluated before the fn(par) is evaluated into
myFxn(parms=par).
A point which is probably, but not certainly, irrelevant: myFxn() has
"..." as its final argument, so as to pass tuning arguments to
integrate(). The function being integrated contains phi(), as well as
other stuff, of a dummy variable. The calls in the debugging tree,
however, are *not* those involving integrate().
It would probably be possible to include some
substitute/eval/as.name/etc. constructions within the function myFxn, in
order to avoid this problem, but as myFxn already involves numeric
integrations whose integrand involves the optimized parameters
themselves, computing on the language at each step of the optimization
seems like a bad idea.
My question: is there a straightforward and efficient way of
constructing this function and optimizing it with optim(), allowing for
the argument phi to pass an arbitrary distribution function whose
parameters are the global ones being optimized? In particular, in the
third environment of the debugger tree, is there a way to force the
fn(par, ...) ----> myFxn(parms=par, ...) evaluation before the "..." get
evaluated?
Thanks, John
John Szumiloski, Ph.D.
Senior Biometrician
Biometrics Research
WP53B-120
Merck Research Laboratories
P.O. Box 0004
West Point, PA 19486-0004
> (215) 652-7346 (PH)
> (215) 993-1835 (FAX)
>
>
#################################################
#### obligatory session info
####
#### Windows XP 2002 sp3, customized by corporate IT
> version
_
platform i386-pc-mingw32
arch i386
os mingw32
system i386, mingw32
status Patched
major 2
minor 9.2
year 2009
month 09
day 05
svn rev 49600
language R
version.string R version 2.9.2 Patched (2009-09-05 r49600)
> search()
[1] ".GlobalEnv" "package:geometry" "package:datasets"
"<mylib1>" "<mylib2>"
[6] "package:VGAM" "package:stats4" "package:Design"
"package:Hmisc" "package:boot"
[11] "package:splines" "package:MASS" "package:nnet"
"package:utils" "package:stats"
[16] "package:graphics" "package:grDevices" "<mylib3>"
"<mylib4>" "<mylib5>"
[21] "<mylib6>" "package:methods" "Autoloads"
"package:base"
Notice: This e-mail message, together with any attachme...{{dropped:15}}
______________________________________________
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.
