Why am I getting inconsistent results with the RandomFields package?
I am intrigued to see if other users have similar results or suggestions as to why this might be happening. I pulled this code from the JSS article here , section 5, and I have different results on different platforms. The algorithm is deterministic (not seed dependent) and I cannot find a pattern as to what is causing the inconsistencies. A bit of background - the purpose of the following lines of code is to estimate parameters in the model using maximum likelihood:
library(RandomFields)
data(weather)
nug <- RMmatrix(M = matrix(nc = 2, c(NA, 0, 0, NA)), RMnugget())
pars.model <- nug + RMbiwm(nudiag = c(NA, NA),
scale = NA,
cdiag = c(NA, NA),
rhored = NA)
Dist.mat <- as.vector(RFearth2dist(weather[, 3:4]))
PT <- weather[, 1:2]
pars <- RFfit(pars.model, distances = Dist.mat, dim = 3, data = PT)
print(pars)
The above result is quite long, but it is sufficient to compare the displayed AIC values, which reflect the optimal result found by the procedure RFfit
. On different platforms, I got two sets of AIC values ββthat are very different from a practical point of view.
CentOS 6 64-bit 64GB RAM Intel (R) Xeon (R) CPU E5-2620 v2 @ 2.10GHz R version 3.2.0 (2015-04-16) x86_64-redhat-linux-gnu (64-bit) AIC : 2547.474
CentOS 6 64-bit RAM 256 GB AMD Opteron (tm) 6376 R 3.2.0 (2015-04-16) x86_64-redhat-linux-gnu (64-bit) processor AIC: 2562.027
Ubuntu 14.04 64-bit RAM 8 GB Intel (R) Celeron (R) CPU N2820 @ 2.13 GHz R version 3.0.2 (2013-09-25) x86_64-pc-linux-gnu (64-bit) AIC: 2547.474
Windows 7 64-bit RAM 8 GB Intel (R) Core (TM) i7-3632QM CPU @ 2.20GHz R version 3.1.2 Patch (2014-11-25 r67060) i386-w64-mingw32 / i386 (32-bit) AIC: 2547.474
Windows 7 64-bit RAM 8 GB Intel (R) Core (TM) i7-3632QM CPU @ 2.20GHz R version 3.1.2 Patch (2014-11-25 r67060) x86_64-w64-mingw32 / x64 (64-bit) AIC: 2562.027
Windows 7 64-bit RAM 8 GB Intel (R) Core (TM) i7-3632QM CPU @ 2.20GHz R version 3.0.2 (2013-09-25) x86_64-w64-mingw32 / x64 (64-bit) AIC : 2562.027
Getting into the internal environment, I found that basic optimization routing is a standard R routine optim
. Are such drastic results expected from any optimizer? Some of them use OpenBLAS, but OpenBLAS is not the reason for this. From the below, I can't figure out if it is OS, R version, package version (they are all the same) or processor architecture ... Any ideas on this are appreciated. Thank.
source to share
No one has answered this question yet
Check out similar questions: