[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [TCML] Re:One factor at a time? was: RSG disks - who can make them ?



 
Hi Ed, Kurt, All,
    
    Yes, multivariate analysis will handle a  multitude of sins. The problem 
is (IMO), that many amateurs and too many  professionals get lost in such 
concepts as non-linear methods, principal  component analysis, and 
partially-correlated variables. The proper design and  interpretation of multivariable 
experiments is beyond the skills of the vast  majority of those all too willing to 
do the experiments.
    Before retiring, I was faced with a problem of  determining the response 
of a series of similar complex systems to several  (actually 8) non-linear 
variables, all of which were out of the observer's  control, and almost all of 
which showed partial correlation. While I had a an  elegant theoretical 
solution, in the 90's it required Cray-level computing  power for reasonable time 
solution, and the "suits with  financial authority" were lost and unconvinced, 
even though it was rated  top paper at an international conference[1].   
    Is anyone familiar with an amateur-affordable  alternative to the SAS(r) 
statistical package including PROC NLIN(r) and  PROC PLOT-3D(r)? While these 
would definitely handle a lot of TC problems, the  cost of ~$3K down and $1.5K 
annual renewal fee is a little stiff for  retired old folks without corporate 
or university privileges.
 
Matt D.
 
[1] http://www.psig.org/Conferences/conf.asp?year=1993
 
In a message dated 11/24/07 7:56:38 A.M. Eastern Standard Time,  
k.schraner@xxxxxxxxxxx writes:

Ed  Phillips wrote:
>    You cannot test anything at all if you  alter more than 1 aspect at
> time!
>    This statement  is worth a comment.  Not many engineers I know have
> learned that  but it's very important if you want to learn
> anything from an  experiment.  One experiment with two variables is
> close to  useless.
> Ed
>
>  _______________________________________________
> Tesla mailing  list
> Tesla@xxxxxxxxxxxxxx
>  http://www.pupman.com/mailman/listinfo/tesla

Hi Ed,

of course, I  agree, especially if evaluating _known_ variables in their 
possibly  nonlinear behavior, ...but not quite full hearted ;-)

During my  jobtime, I had sometimes to do with the design and analysis of 
experiments  for the chemical industry. Sometimes quite a number of 
variables(5..to  10?) were candidates for a response-function, and it was not 
known, which  ones were significant...and which one might interact with one 
another. In  those situations multivariate statistical analysis, with the 
goal of  reducing the experimental work/cost is attained by varying different 
 
variables _simultaneously_, following statistically founded experimental  
design plans for the different levels of detail. Interpretation models at  
those levels are predominately simple linear or second order  relationships. 
But, if the levels of the variables are chosen by good  common sense, 
results can be convincing.

Literature: Design and  Analysis of Industrial Experiments
Owen L. Davies, Publ. by Oliver & Boyd, London, 1967

In  electricity we are much more used to strong physics-law relationships,  
than in chemical engineering. But a TC, despite it's few elements,  presents 
a not quite easy to understand multivariable system, as i.e. is  shown by the 
current thread of Ch.Swinson/J.Freau/B.Anderson. Part of it  might be solved 
by the, hopefully not forgotten, Terry  program:

http://drsstc.com/~sisg/files/scantesla/scantesla762.zip

...which  tries to exhaust a wide variable space, by applying the known 
physics  laws.

Best regards

Kurt







**************************************Check out AOL's list of 2007's hottest 
products.
(http://money.aol.com/special/hot-products-2007?NCID=aoltop00030000000001)
_______________________________________________
Tesla mailing list
Tesla@xxxxxxxxxxxxxx
http://www.pupman.com/mailman/listinfo/tesla