In fact I have been in fairly regular communication with Ryan Nielson
about this, but I don't get the impression he has a lot of time to
devote to resolving it. So far I have managed to circumvent all the
potential problems he has suggested, and I have narrowed it down to
this Fortran loop. Now he seems of the mind that it is a problem with
R - which I suppose is possible? - but whether it is the Fortran code,
R, or something else I don't believe I have the tools to assess the
problem any further. I am certainly not imagining the problem, as
I've mentioned, but if somebody more knowledgeable could at least
reproduce it we might be somewhere closer to a solution. In the
meantime I have suggested that BBMM be used with caution, especially
in big loops with large data sets, but Ryan hasn't commented on this,
and at any rate I wouldn't mind getting a third or fourth opinion.
Tyler
On 2010-07-14, at 5:33, Roger Bivand <roger.biv...@nhh.no> wrote:
On Tue, 13 Jul 2010, Tyler Dean Rudolph wrote:
I've been constructing utilization distributions for 63 animal
movement
trajectories using a Brownian bridge movement model with BBMM. I
started
noticing after running the routine numerous times at different
resolutions
that certain UDs would be grossly wrong, meaning instead of
returning a
strongly negative exponential distribution of probabilities (UD
$probability)
that sums to 1, there might be two big modes in opposite places, or
negative
probability values, or simply "NaN" across the board. I have found
this
problem on Mac, Linux, and Windows machines in various versions of R
(including the most recent), and I was hoping someone in the R
community
would be able to verify this and perhaps speculate on why this is
occurring.
Have you also contacted the package authors directly? They would be
the best point of entry here, I think.
Roger
The problem appears to originate either 1) in the FORTRAN code that
estimates BMVar and calculates the BBMM, or 2) in the way R handles
its
operation. I've broken down BBMM into its essential parts, and in
the
attached environment are all of the objects necessary to run the
FORTRAN
code. I have set it up in a loop here so it continually estimates
a UD on
the same data then and produces a histogram, which I have found to
be most
revealing in identifying problems. If the code stops running it's
because
the result was all "NaN". If you run the loop repeatedly you may
find, as I
have, that while the results may initially appear more-or-less
consistent,
at other moments they will change, sometimes gradually, and sometimes
quickly. If you are looking at it and it is not changing, I
encourage you
to remove all objects, close and reopen R, reload the workspace and
run the
same loop again. If you do this numerous times you should
eventually see
unusual results, meaning inconsistent estimation, or wacky
probability
distributions, or "NaN".
###
###
###
###
####################################################################
## Run the Fortran routine repeatedly with identical data
## - all required objects are in workspace "BBMM_BUG.RData" -
require(BBMM)
par(mfrow=c(2,2))
repeat {
ans <- .Fortran("BBMM", as.integer(n.locs), as.integer
(grid.size),
as.double(time.lag), as.double(T.Total), as.double(x),
as.double(y), as.double(BMvar), as.double(location.error),
as.double(area.grid$x), as.double(area.grid$y),
as.double(time.step),
as.double(probability), PACKAGE = "BBMM")
if(!sum(ans[[12]])=="NaN") { hist(ans[[12]]) } else { break }
}
###
###
###
###
#####################################################################
Tyler
--
Roger Bivand
Economic Geography Section, Department of Economics, Norwegian
School of
Economics and Business Administration, Helleveien 30, N-5045 Bergen,
Norway. voice: +47 55 95 93 55; fax +47 55 95 95 43
e-mail: roger.biv...@nhh.no
_______________________________________________
R-sig-Geo mailing list
R-sig-Geo@stat.math.ethz.ch
https://stat.ethz.ch/mailman/listinfo/r-sig-geo
