[R] Dynamic Memory Allocation Errors in 1.2.0?

[Kurt Hornik]

>>>>> Jeff Lischer writes:

> Let me first say I'm new to R and new to this mail list, so I apologize
> ahead if this issue has already been discussed here.

> The dynamic memory allocation of 1.2.0 seems to be an improvement over the
> static allocation of 1.1.1. However, I have run across at least one case
> where the dynamic allocation might be corrupting my results. I am
> interested in calculating the 95% confidence margins of performance results
> based on the number of wins, draws, and losses in a match. For example, for
> a match with 10 wins, 20 draws, and 10 losses, the observed performance is
> 50% and the 95% margins are about (39.0%, 61.0%).

> I wrote the following simple file to calculate these margins using the ABC
> method from Resampling:

>      # main function to calculate margins based on ABC method
>      margins.abc <- function(W,D,L) {
>         # Create arrays with correct number of wins, draws, and losses
>         w <- rep(1.0,W)
>         d <- rep(0.5,D)
>         l <- rep(0.0,L)

>         # Call abcnon with the mean as the function being used
>         results.abc <<- abcnon(c(w,d,l),function(p,x) {sum(p*x)/sum(p)},
>                         alpha = c(0.025, 0.975))

>         print("Actual Results")
>         print(c(results.abc$stats$t0))
>         print("Estimated 95% Margins")
>         results.abc$limits
>      }

> The 'abcnon' routine is from the 'bootstrap' package and it also calculates
> the margins based on standard theory so you can easily see the difference
> between standard and ABC margins. In R 1.1.1, my file always seems to work
> just fine. However, when I run the same file in 1.2.0 I get incorrect
> results when the number of games in the match is large.

> ** Score **    **************  95 % Margins (%)   *************
> W/D/L          Standard  1.1.1           1.2.0

> 10/20/10  (39.0, 61.0)   (39.0, 61.0)    (39.0, 61.0)
> 100/200/100    (46.5, 53.5)   (46.5, 53.5)    (46.5, 53.5)
> 400/800/400    (48.3, 51.7)   (48.3, 51.7)    (45.8, 49.2)
> 800/1600/800   (48.8, 51.2)   (48.8, 51.2)    (52.8, 55.8)

> The 1.1.1 results are always the same as the standard results, which
> is correct for these symmetric cases. The 1.2.0 results, however, are
> meaningless for the largest two cases as they don't even enclose the
> observed result of 50%. My guess is that some critical data is getting
> written over by the dynamic memory allocation. I got the same results
> on Windows 98 and Windows NT with various settings of the
> min/max/vsize/nsize parameters. I also tried using the 'abc.ci'
> routine from the 'boot' package but got similar results for large
> matches.

> Has anyone else observed this behaviour in 1.2.0? Is there anything I
> can do differently to prevent it from happening? Until I figure this
> out, I will stay with 1.1.1 -- I don't want to be always worrying
> about the validity of the 1.2.0 results.

On Debian GNU/Linux, I get

R> margins.abc(400, 800, 400)
[1] "Actual Results"
[1] 0.5
[1] "Estimated 95% Margins"
     alpha       abc      stan
[1,] 0.025 0.4826762 0.4826762
[2,] 0.975 0.5173238 0.5173238
R> margins.abc(800, 1600, 800)
[1] "Actual Results"
[1] 0.5
[1] "Estimated 95% Margins"
     alpha       abc      stan
[1,] 0.025 0.4877502 0.4877502
[2,] 0.975 0.5122498 0.5122498

which seems right.  Maybe this is only a problem on Windows?

-k
-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-help mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !)  To: r-help-request at stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._