From: <HIDDEN>
Newsgroups: comp.soft-sys.matlab
Subject: Re: creating loops for a large system of variables
Date: Mon, 04 Apr 2011 09:48:42 -0500
Organization: NNTP Server
Lines: 52
Message-ID: <incloi$ued$>
References: <ina9ds$qcn$> <inab9s$o34$> <inaj1d$j5b$> <inb174$n3k$> <inb3v8$4n6$> <inb7hd$rku$> <inba0a$5nl$> <inbadn$rfb$> <inbaqn$hkv$> <inbbsc$3h4$> <inbccs$uui$> <inbggf$e2g$> <inbhnq$9q3$>
Mime-Version: 1.0
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
User-Agent: Mozilla/5.0 (Windows; U; Windows NT 5.1; en-US; rv: Gecko/20101207 Thunderbird/3.1.7
X-Notice: Filtered by postfilter v. 0.8.2
Xref: comp.soft-sys.matlab:719957

On 4/3/2011 11:34 PM, dpb wrote:
> On 4/3/2011 11:13 PM, Nate Jensen wrote:
>>> Or "Latin-Hypercube Sampling" is another buzzword
>>> More actual detail of what is needed could undoubtedly help cull the
>>> general to at least a little more specific.
>>> --
>> What I need to do is understand the behavior of my system as different
>> design variables are changed within it.
>> Although I could run each case independently, it would be to my
>> advantage to compare them to one another in one run. At the moment,
>> optimization or inverse design of the inputs, would be much more
>> difficult, due to their complex, and at times rigid, nature. The main
>> caveat that I have is that, many of the inputs depend upon one another,
>> and may or may not be changed depending upon the design. This all sums
>> up to the ability to change various inputs depending on the design of
>> the system, and compare their outputs.

>> I am regretful that I am not making myself clear.
> No, that's pretty clear. Read up on response surfaces as one possible
> tool--if the models aren't terribly nonlinear. Steven L's pointers are
> in similar direction.
> DOE is the exercise in setting up criteria based on combination of what
> is known of the model and the desired estimators to try to help in an
> efficient use of the number of points and values for the independent
> variables of those points that meet specific design criteria regarding
> which effects may be estimated w/o confounding, which effects are
> possibly sacrificed or confounded w/ others to reduce the design size,
> etc., etc., ...
> If this is academic exercise, if you have access to a statistician it
> would be _a_good_thing_ (tm)....

And to make Steve's and my point explicit...if you have 25 variables to 
study and you only use two levels for each, that's 2^25 total cases in 
the simplistic approach or 33554432 runs of the model.  If there's any 
nonlinearity at all in the model you need a minimum of three levels to 
see it or 3^25 ~ 8.5E11.  I don't think you'd be able to interpret the 
results even if you could run all those cases.

So, you need a smarter approach.