Jörgen Ullberg wrote:
>> There were recently mensioned in an answer to someone who needed to do
> some kind of ordination technique?, a software called KYST. In the
> answer it was said that KYST could be obtained from any netlib archive.
> I am not that fluent in computer lingo so what does netlib archive stand
> for? I also tried some searching on the internet for KYST without
> success. Anyone having a more specific web-address for KYST?
>Since you are from .se, the nearest netlib archive to you is in Oslo at
http://www.netlib.no/. It is a mirror site of
http://netlib.bell-labs.com/netlib/. Following links, you can find a
list of other mirror sites.
KYST is the mother of non-metric multidimensional scaling programs,
named after Kruskal, Young, Torgeson & Shepard (but pronounced
`kissed'). Last time I tried, I could compile it with f2c + gcc (but not
directly with g77). However, it is old-time Fortran (or, I should write,
FORTRAN), and you may get plenty of warnings of initializing
non-character variables with character strings (it is that old Fortran)
etc. Moreover, you do need the manual or you just don't know how to run
the program, but you can get that from netlib as well. You may run in
the trouble with large data sets (>1800 distances, say), and changing
the dimensions is not quite straightforward in KYST due to awkward,
Nowadays KYST pedigrees or their alternative, ALSCAL pedigrees, are
available in many standard statistical packages and therefore you are
not dependent on KYST. Just check your local statistical program. In
particular, the MASS package of Venables & Ripley for R + S-plus has
Kruskal scalling (isoMDS) and, R+ S-plus have other MDS variants as
well, up to Sammon scaling and principal curves. However, my mate in
Utah (Dave Roberts) is dissatisfied with tie-handling in MASS isoMDS.
The only MDS program that really tries to avoid trapping into local
minima with multiple random starts and Procrustes comparison of results
is Peter Minchin's DECODA which uses a tested and effective variant of
KYST. I have heard that version 4 of PC-Ord does almost the same, but I
haven't seen this version (although they once promised to send me a test
copy). However, both are proprietary programs, and they may be ill
suited for your needs if you are not in their particular field of
research (vegetation ecology). Many statistical packages use metric
scaling or PCA as a starting solution which is almost guaranteed to lead
always in a global minimum far away from the global minimum (since the
metric solution quite often is very close to a local minimum). I have
heard rumours of many attempts of using simulated annealing in criterion
minimization, but I have not heard of any finished programs.
If you do not have S-plus or money to get S-plus, you can get R for your
platform at http://cran.r-project.org/ which lists several mirror sites
(one, at least, in Denmark).
J.Oksanen -- UHLA (University of Helsinki at Lahti)
C-h i m emacs RET m text RET m formatted RET
"GNU Emacs is evil." Linus Torvalds