R help - Aug 2008 - ecdf manipulation

I’m using the edcf function to look at a number of empirical distributions
graphically for run-time analyses of stochastic optimization algorithms.
When dealing with problems where the optimal solution for these problems is
always found everything is fine and the graphs are very useful for
comparative observations. These distributions have a vertical axis height of
one i.e. a probability of one. However, I’ve hit a problem when the optimal
solution is not always obtained during the allotted run-time. In the cases
I’m looking these graphs are only concerned with the behaviour those runs
that find the optimal solution.
 
e.g. say we have two algorithms one solves a given problem 1000 times out of
1000 runs and the second solves the same problem 800 times out of 1000 runs
then the first plot rises from 0 to 1 where as the second should only rise
to 0.8
 
One idea is that the ecdf R code relies upon the number of samples n (1000
in this case) is it possible to manipulate this R code and pass an extra
argument to have n defined when the function is called, opposed to the value
of n being set to the size of the vector being passed in as appears to be
the current case, whilst maintaining its graphical capability?
 
If so how and where do I get hold of the ecdf R code to manipulate?
If not then does anyone have any suggestions?
 
Thanks
 
Harry Venables

Internal Virus Database is out-of-date.
Checked by AVG. 

05:48
 

	[[alternative HTML version deleted]]

>>>>>   <harryandlaura at talktalk.net>
>>>>>     on Fri, 15 Aug 2008 00:51:56 +0100 writes:
    > I?m using the edcf function to look at a number of
    > empirical distributions graphically for run-time analyses
    > of stochastic optimization algorithms.  When dealing with
    > problems where the optimal solution for these problems is
    > always found everything is fine and the graphs are very
    > useful for comparative observations. These distributions
    > have a vertical axis height of one i.e. a probability of
    > one. However, I?ve hit a problem when the optimal
    > solution is not always obtained during the allotted
    > run-time. In the cases I?m looking these graphs are
    > only concerned with the behaviour those runs that find the
    > optimal solution.
 
    > e.g. say we have two algorithms one solves a given problem
    > 1000 times out of 1000 runs and the second solves the same
    > problem 800 times out of 1000 runs then the first plot
    > rises from 0 to 1 where as the second should only rise to
    > 0.8
 
    > One idea is that the ecdf R code relies upon the number of
    > samples n (1000 in this case) is it possible to manipulate
    > this R code and pass an extra argument to have n defined
    > when the function is called, opposed to the value of n
    > being set to the size of the vector being passed in as
    > appears to be the current case, whilst maintaining its
    > graphical capability?

Yes, this is possible 
 { install.packages("fortunes"); fortunes::fortune("Yoda") }
 
    > If so how and where do I get hold of the ecdf R code to
    > manipulate?

  https::/svn.r-project.org/R/trunk/src/library/stats/R/ecdf.R

always has the (R-devel version of the) ecdf code;
you may also want to study the  stepfun code in stepfun.R (same
directory)  which is made use of by ecdf and its methods.

Martin Maechler,
ETH Zurich


    > If not then does anyone have any suggestions?
 
    > Thanks
    > Harry Venables