R help - Feb 2003 - (no subject)

Dear R helpers,

I have a curious problem, which is that a program I have written in R
crashes R, unpredictably. When I say that the program crashes R, I mean
that it causes R to terminate, completely. But, the operating system
(Windows) continues OK.

The program loops around a randomForest regression. I am trying to use
randomForest to predict the diagnosis for a set of patients with dementia.
I have 140 patients, of whom 125 have one kind of dementia and 15 have
a different kind. We have measured about 50 variables for each patient.
The program repeatedly and randomly splits the 150 patients into 75:25 
groupings. Then it 'trains' the randomForest using the 75% grouping,
then
it 'predicts' the remaining 25% of cases. I save the predictions and at
the end compute a Receiver Operating Characteristic curve.

The program will work away happily for hours and will sometimes perform
as many as 100 cycles around the above loop. It gives quite nice results.
But, quite frequently, and unpredictably, it fails and R just stops. As I 
say, the operating system continues and any other programs that are multi-
tasking are unaffected.

If anyone could give me an idea on how to avoid this problem, I'd be
grateful.

Thanks,

Jonathan Williams


Jonathan Williams
OPTIMA
Radcliffe Infirmary
Woodstock Road
OXFORD OX2 6HE
Tel +1865 (2)24356

On Thu, 13 Feb 2003, Jonathan Williams wrote:
> The program will work away happily for hours and will sometimes perform
> as many as 100 cycles around the above loop. It gives quite nice results.
> But, quite frequently, and unpredictably, it fails and R just stops. As I
> say, the operating system continues and any other programs that are multi-
> tasking are unaffected.
>
> If anyone could give me an idea on how to avoid this problem, I'd be
> grateful.
It sounds as though randomForest has a small bug in its C memory
management.  Unfortunately these can be quite hard to find and fix and
there isn't any straightforward way to work around them.

	-thomas

Thomas Lumley			Asst. Professor, Biostatistics
tlumley at u.washington.edu	University of Washington, Seattle