s.kcv <contrib>	      GRASS Reference Manual	  <contrib> s.kcv



NAME
     s.kcv - Randomly partition sites into test/train sets.
     (GRASS Sites Program)

SYNOPSIS
     s.kcv
     s.kcv help
     s.kcv [-dq] k=value sites=name

DESCRIPTION
     s.kcv randomly divides a sites lists into k sets of
     test/train data (for k-fold cross validation).  Test
     partitions are mutually exclusive. That is, a site will
     appear in only one test partition and k-1 training
     partitions.

     The program generates a random point using the selected
     random number generator and then finds the closest site to
     it. This site is removed from the candidate list (meaning
     that it will not be selected for any other test set) and
     saved in the first test partition file.  This is repeated
     until enough points have been selected for the test
     partition.	 The number of sites chosen for test partitions
     depends upon the number of sites available and the number of
     partitions chosen (this number is made as consistent as
     possible while ensuring that all sites will be chosen for
     testing). This process of filling up a test partition is
     done k times.

     Flags:

     -d		       Use drand48() (default is rand()).

     -q		       Quiet. Don't report progress.

     Parameters:

     k=value	       Positive integer value indicating the
		       number of partitions.

     sites=name	       Name of a sites file to store random
		       points in.

     Test/train pairs are saved as sites list using name as a
     basename. Test sites are saved in name-test.i while training
     sites are saved in name-train.i, where i ranges from zero to
     k.

NOTES
     Existing files are silently overwritten.





GRASS 5.0beta7	      GRASS Development Team			1






s.kcv <contrib>	      GRASS Reference Manual	  <contrib> s.kcv



     An ideal random sites generator will follow a Poisson
     distribution. The merits of rand() versus drand48() in this
     respect are unknown.  Train/test partitions, though, can
     only be as random as the original sites.  This program
     simply divides sites up in a random manner.

     Be warned that random number generation occurs over the
     intervals defined by the current region.

     This program may not work properly with Lat-long data.

SEE ALSO
     rand(3), drand48(3), s.rand and g.region

BUGS
     Please send all bug fixes and comments to the author.

AUTHOR
     James Darrell McCauley, Purdue University
     (mccauley@ecn.purdue.edu)

NOTICE
     This program is part of the contrib section of the GRASS
     distribution.  As such, it is externally contributed code
     that has not been examined or tested by the Office of GRASS
     Integration.





























2		      GRASS Development Team	   GRASS 5.0beta7