[Rd] different results on linux and windows

[Paul Gilbert]

I cannot say what the problem is in your code, but in general it is 
possible to get the same random sequences from Linux and Windows with 
R's Mersenne Twister generator.  If you generate long sequences (say 10s 
of thousands) and then start doing comparisons involving remainders, 
like differencing the sums of two mean zero sequences, then differences 
between libraries will show up. I have found bigger differences between 
Linux variants than I have between Windows and most Linuxes.

Paul Gilbert

Klaus Nordhausen wrote:
> Dear R experts,
> 
> we are preparing an R-package to compute the Oja Median which contains
> some C++ code in which random numbers are needed. To generate the random
> numbers we use the following Mersenne-Twister implementation:
> 
> // MersenneTwister.h
> // Mersenne Twister random number generator -- a C++ class MTRand
> // Based on code by Makoto Matsumoto, Takuji Nishimura, and Shawn Cokus
> // Richard J. Wagner  v1.0  15 May 2003  rjwagner at writeme.com
> 
> the random seed for the Mersenne-Twister is provided by our R-function
> which gives an (random) integer to  the C++ function srand() which in
> turn sets the seed in the code.
> 
> Using the set.seed in R makes now the results reproducible, but the
> results differ between windows and linux.
> 
> Does anyone know what the problem there is?
> 
> Our suspicion is that the reason is that some libraries are different
> implemented on linux and windows (XP) compilers.
> 
> After the program start we set the seed in row 447(vkm.cpp) with 
> srand(int);
> 
> When the median will be calculated, an intern seed is set with unsigned
>  int seed = rand();   ( in row 100 (vkm.cpp)). This seed will be used
> to calculate some random subsets and to
> create a Mersenne Twister object with MTRand rr(seed);  (row 156, vkm.cpp).
> 
> The MTRand Object rr is called with an unsigned Integer, so the
> important function in the mersenneTwister.h class is in line 87:
> MTRand( const uint32& oneSeed );
> 
> According to that the Random Number Generator uses the methods
> initialize(oneSeed); and reload();  (inside the method, beginning in
> line 215)
> 
> This both methods (line 283 and line 301) are using beside others
> registers. Could it be that there is a different behavior between
> Windows and Linux?
> 
> We do not want to use only srand() since we might need more then the
> number of pseudo random numbers that algorithm can provide.
> 
> For those interested and which would like to see the code, a first
> version of the package, called OjaMedian, is available as source file
> and windows binary on my homepage:
> http://www.uta.fi/~klaus.nordhausen/down.html
> 
> The problem is in the ojaMedian function when the evolutionary algorithm
> is used. Involved C++-files are mainly vkm.cpp and MersenneTwister.h.
> 
> We would be very grateful for any advice on how to solve this problem.
> (below is also a demonstration)
> 
> Thank you very much in advance,
> 
> Klaus
> 
> Results on windows XP:
> 
> Compiler used: gcc version 4.2.1-sjlj (mingw32-2)
> 
>> library(OjaMedian)
>> set.seed(1)
>> testD <- rmvnorm(20,c(0,0))
>> summary(testD)
>        V1                V2
>  Min.   :-2.2147   Min.   :-1.989352
>  1st Qu.:-0.3844   1st Qu.:-0.399466
>  Median : 0.3597   Median :-0.054967
>  Mean   : 0.1905   Mean   :-0.006472
>  3rd Qu.: 0.7590   3rd Qu.: 0.655663
>  Max.   : 1.5953   Max.   : 1.358680
>> set.seed(1)
>> ojaMedian(testD)
> [1]  0.21423705 -0.05799643
>> sessionInfo()
> R version 2.9.0 (2009-04-17)
> i386-pc-mingw32
> 
> locale:
> LC_COLLATE=Finnish_Finland.1252;LC_CTYPE=Finnish_Finland.1252;LC_MONETARY=Finnish_Finland.1252;LC_NUMERIC=C;LC_TIME=Finnish_Finland.1252 
> 
> 
> attached base packages:
> [1] stats     graphics  grDevices utils     datasets  methods   base
> 
> other attached packages:
> [1] OjaMedian_0.0-14 ICSNP_1.0-3      ICS_1.2-1        survey_3.14
> [5] mvtnorm_0.9-5
> 
> loaded via a namespace (and not attached):
> [1] tools_2.9.0
>>
> 
> Results on Linux Kubuntu 8.10
> result of: cat /proc/version:
> Linux version 2.6.28-11-generic (buildd at palmer) (gcc version 4.3.3
> (Ubuntu 4.3.3-5ubuntu4) ) #42-Ubuntu SMP Fri Apr 17 01:57:59 UTC 2009
> 
>> library(OjaMedian)
>>  set.seed(1)
>>  testD <- rmvnorm(20,c(0,0))
>>  summary(testD)
> 
>        V1                V2
>  Min.   :-2.2147   Min.   :-1.989352
>  1st Qu.:-0.3844   1st Qu.:-0.399466
>  Median : 0.3597   Median :-0.054967
>  Mean   : 0.1905   Mean   :-0.006472
>  3rd Qu.: 0.7590   3rd Qu.: 0.655663
>  Max.   : 1.5953   Max.   : 1.358680
> 
>> set.seed(1)
>>  ojaMedian(testD)
> 
> (-0.501381, 0.193929)[1] 0.119149071 0.002732100
> 
>> sessionInfo()
> 
> R version 2.8.1 (2008-12-22)
> i486-pc-linux-gnu
> 
> locale:
> LC_CTYPE=en_US.UTF-8;LC_NUMERIC=C;LC_TIME=en_US.UTF-8;LC_COLLATE=en_US.UTF-8;LC_MONETARY=C;LC_MESSAGES=en_US.UTF-8;LC_PAPER=en_US.UTF-8;LC_NAME=C;LC_ADDRESS=C;LC_TELEPHONE=C;LC_MEASUREMENT=en_US.UTF-8;LC_IDENTIFICATION=C 
> 
> 
> attached base packages:
> [1] stats     graphics  grDevices utils     datasets  methods   base
> 
> other attached packages:
> [1] OjaMedian_0.0-14 ICSNP_1.0-3      ICS_1.2-1        survey_3.14
> [5] mvtnorm_0.9-5
> 
> 
> 
> 
> 
====================================================================================

La version française suit le texte anglais.

------------------------------------------------------------------------------------

This email may contain privileged and/or confidential in...{{dropped:26}}