Re: [office] Re: [office-comment] Re: sigmoid curve fitting in trend lines

[robert_weir@us.ibm.com]

office@lists.oasis-open.org wrote on 11/05/2012 03:19:56
PM:

> From: Andreas J Guelzow <andreas.guelzow@concordia.ab.ca>
> To: office@lists.oasis-open.org, 
> Date: 11/05/2012 03:20 PM
> Subject: Re: [office] Re: [office-comment] Re:
sigmoid curve fitting
> in trend lines
> Sent by: office@lists.oasis-open.org
> 
> Well, personally I see little point in specifying an attribute value
> that will not lead to any interoperability. In that case I think it
is
> much better to allow name-spaced attribute values so that different
> implementations can easily add their favourite best-fit lines, or
at
> least only add attribute values if at least one implementation is
> willing to implement it and describe its algorithm.
> 
> Just adding an attribute value nobody supports seems to be pointless.
> 

Well, that is a different issue, a good one,
but different.

Two questions:

1) Does it make sense to specify requirements
for numeric calculations that are not guaranteed to yield identical results?
 I'd say yes, and we already do this.  OpenFormula does not mandate
a particular floating point processing model, or precision.  So even
basic arithmetic operations are not guaranteed to be identical, although
in practice most implementations will.  Even linear regression, if
you have multicollinearity issues, can yield solutions that are sensitive
to implementation details beyond what we currently specify.  I think
this is fine -- a fact of life with numeric calculations -- and extending
this to non-linear models is OK.

2) Is anyone going to implement it?  
I don't think that we want the ODF spec to be a preferred way for users
to lobby for product features that they want their vendors to add.  That
would be backwards.   So If no one is planning to implement this then
I don't see any good reason to add it.  And as I've said previously,
an even better indicator would be if more than one implementor is planning
to support it.

-Rob

> Andreas
> 
> 
> On Mon, 2012-11-05 at 13:04 -0700, robert_weir@us.ibm.com wrote:
> > office@lists.oasis-open.org wrote on 10/26/2012 12:32:24 AM:
> > 
> > > From: "Andreas J. Guelzow" <andreas.guelzow@concordia.ab.ca>
> > > To: office@lists.oasis-open.org, 
> > > Date: 10/26/2012 12:32 AM 
> > > Subject: Re: [office] Re: [office-comment] Re: sigmoid curve
fitting
> > > in trend lines 
> > > Sent by: office@lists.oasis-open.org 
> > > 
> > > On Thu, 2012-10-25 at 15:57 -0600, robert_weir@us.ibm.com
wrote:
> > > > I'm not sure we want to specify the method used to
fit the curve.
> >  We
> > > > should probably specify the outcome.   The specific
algorithm used
> > > > might vary according to the processor, number of processors,
etc. 
> > > > 
> > > > I'm sure Andreas would have some ideas, but I wonder
if we could
> > say
> > > > something like, "Uses an implementation-defined
method to fit a
> > > > sigmoid curve to the data while minimizing the average
squared
> > > > deviation".  Then we're specifying the outcome,
but allowing fo
> > > > innovation in the method.
> > > 
> > > 
> > > Are we sure that the proposals (and the calculation supposedly
> > outlined
> > > in those papers I haven't read,) in fact create a curve
that
> > minimizes
> > > the average squared deviation?
> > > 
> > > I am not used to R, but looking at the example referenced
in a
> > previous
> > > message in this thread and considering R's SSlogis function
it seems
> > to
> > > me that R's fitting depends not only on the data but also
on three
> > > initial values. Are we sure that there is a unique sigmoid
function
> > that
> > > can be fitted to a given set of data?
> > > 
> > > If there is more than one than a generic description as
above will
> > not
> > > provide for interoperability.
> > > 
> > 
> > So what are our choices? 
> > 
> > 1) Say something in very general terms, like: "attempts
a 'best fit'
> > of the data to the following model:" This allows implementations
to
> > vary on the quality of the fit, but doesn't guarantee identical
> > results.  So similar to IRR(). 
> > 
> > 2)Cite a specific algorithm, with the risk that it might not
be the
> > best one today, or the best one 5 years from now.  But it
will give
> > the same results to the extent everyone implements the cited
> > algorithm. 
> > 
> > IIRC once you get beyond linear models, or ones that can be
> > transformed to linear models, you no longer generally have simple
> > techniques that easily yield the same answer.  You have
optimization
> > heuristics that will depend not only on parameters of the mode,
but
> > also parameters of the algorithm, like step size, number of
> > iterations, etc.  So although in theory we could do #2,
wouldn't we be
> > 2nd guessing the implementor on a lot of these details? 
> > 
> > -Rob 
> > 
> > > Andreas
> > > 
> > > 
> > > > 
> > > > 
> > > > Patrick Durusau <patrick@durusau.net> wrote on
10/25/2012 05:26:46
> > PM:
> > > > 
> > > > > From: Patrick Durusau <patrick@durusau.net>
> > > > > To: James Salsman <jsalsman@gmail.com>,
> > > > > Cc: office-comment@lists.oasis-open.org 
> > > > > Date: 10/25/2012 05:26 PM 
> > > > > Subject: Re: [office-comment] Re: sigmoid curve
fitting in trend
> > > > lines 
> > > > > 
> > > > > James,
> > > > > 
> > > > > First, thanks for the references!
> > > > > 
> > > > > Second, standards can refer to other standards
but generally not
> > to 
> > > > > periodical literature as normative.
> > > > > 
> > > > > Which means we will have to extract the necessary
definitions
> > and 
> > > > > incorporate them in the text.
> > > > > 
> > > > > I will grab copies of them in the next day or
so to see what
> > that
> > > > will take.
> > > > > 
> > > > > Hope you are having a great week!
> > > > > 
> > > > > Patrick
> > > > > 
> > > > > On 10/24/2012 09:43 PM, James Salsman wrote:
> > > > > > Patrick,
> > > > > >
> > > > > > Thank you for your question.
> > > > > >
> > > > > > First, please change "Gompertz sigmoid
– Regression with a
> > > > > > logistic...." to "Gompertz sigmoid
– Regression with a
> > > > Gompertz...."
> > > > > > on the fifth line of my original request
below.  I'm sorry
> > about
> > > > that
> > > > > > typo, and I tried to fix it, but it apparently
didn't get
> > fixed in
> > > > my
> > > > > > request.
> > > > > >
> > > > > > As for the methods of calculating the parameters
for fitting
> > those
> > > > two
> > > > > > very important curves, instead of trying
to describe them,
> > could
> > > > you
> > > > > > please reference these authoritative descriptions
of the
> > > > algorithms
> > > > > > instead?
> > > > > >
> > > > > > For logistic curve fitting: J. A. Nelder
(March 1961) "The
> > Fitting
> > > > of
> > > > > > a Generalization of the Logistic Curve"
_Biometrics_ vol. 17,
> > no.
> > > > 1,
> > > > > > pp. 89-110: http://www.jstor.org/stable/2527498
> > > > > >
> > > > > > For Gompertz curve fitting: Karl W. Kaufmann
(1981) "Fitting
> > and
> > > > using
> > > > > > growth curves" _Oecologia_ vol. 49,
no. 3, pp. 293-299:
> > > > > > http://www.springerlink.com/content/w153v81q72400q62
> > > > > >
> > > > > > Please let me know if I can help any further.
> > > > > >
> > > > > > Best regards,
> > > > > > James Salsman
> > > > > >
> > > > > >
> > > > > > On Wed, Oct 24, 2012 at 3:28 PM, Patrick
Durusau 
> > > > > <patrick@durusau.net> wrote:
> > > > > >> James,
> > > > > >>
> > > > > >> With regard to your request for the addition
of:
> > > > > >>
> > > > > >>> "* logistic sigmoid – Regression
with a logistic sigmoid
> > > > function –
> > > > > >>> approximate the values of the series
using the model: y =
> > > > > >>> A+B/(1+e^(-(x-C)/D))
> > > > > >>>
> > > > > >>> "* Gompertz sigmoid – Regression
with a logistic sigmoid
> > > > function –
> > > > > >>> approximate the values of the series
using the model: y =
> > > > > >>> A+B·e^(C·e^(D·x))"
> > > > > >>
> > > > > >> To the ODF standard.
> > > > > >>
> > > > > >> Can you define the method for calculating
the parameters of
> > these
> > > > two
> > > > > >> functions?
> > > > > >>
> > > > > >> Please pardon our delay in reaching your
request.
> > > > > >>
> > > > > >> The TC has not made a decision about
your request but having
> > the
> > > > additional
> > > > > >> information will make that task easier.
> > > > > >>
> > > > > >> Thanks!
> > > > > >>
> > > > > >> Hope you are having a great week!
> > > > > >>
> > > > > >> Patrick
> > > > > >>
> > > > > >> --
> > > > > >> Patrick Durusau
> > > > > >> patrick@durusau.net
> > > > > >> Former Chair, V1 - US TAG to JTC 1/SC
34
> > > > > >> Convener, JTC 1/SC 34/WG 3 (Topic Maps)
> > > > > >> Editor, OpenDocument Format TC (OASIS),
Project Editor
> > ISO/IEC
> > > > 26300
> > > > > >> Co-Editor, ISO/IEC 13250-1, 13250-5 (Topic
Maps)
> > > > > >>
> > > > > >> Another Word For It (blog): http://tm.durusau.net
> > > > > >> Homepage: http://www.durusau.net
> > > > > >> Twitter: patrickDurusau
> > > > > >>
> > > > > 
> > > > > -- 
> > > > > Patrick Durusau
> > > > > patrick@durusau.net
> > > > > Former Chair, V1 - US TAG to JTC 1/SC 34
> > > > > Convener, JTC 1/SC 34/WG 3 (Topic Maps)
> > > > > Editor, OpenDocument Format TC (OASIS), Project
Editor ISO/IEC
> > 26300
> > > > > Co-Editor, ISO/IEC 13250-1, 13250-5 (Topic Maps)
> > > > > 
> > > > > Another Word For It (blog): http://tm.durusau.net
> > > > > Homepage: http://www.durusau.net
> > > > > Twitter: patrickDurusau
> > > > > 
> > > > > 
> > > > > -- 
> > > > > This publicly archived list offers a means to
provide input to
> > the
> > > > > OASIS Open Document Format for Office Applications
> > (OpenDocument)
> > > > TC.
> > > > > 
> > > > > In order to verify user consent to the Feedback
License terms
> > and
> > > > > to minimize spam in the list archive, subscription
is required
> > > > > before posting.
> > > > > 
> > > > > Subscribe: office-comment-subscribe@lists.oasis-open.org
> > > > > Unsubscribe: office-comment-unsubscribe@lists.oasis-open.org
> > > > > List help: office-comment-help@lists.oasis-open.org
> > > > > List archive:
> > http://lists.oasis-open.org/archives/office-comment/
> > > > > Feedback License:
> > > > http://www.oasis-open.org/who/ipr/feedback_license.pdf
> > > > > List Guidelines:
> > http://www.oasis-open.org/maillists/guidelines.php
> > > > > Committee:
> > > > http://www.oasis-open.org/committees/tc_home.php?wg_abbrev=office
> > > > > Join OASIS: http://www.oasis-open.org/join/
> > > > > 
> > > 
> > > -- 
> > > Andreas J. Guelzow, PhD, FTICA
> > > Concordia University College of Alberta
> > > [attachment "signature.asc" deleted by Robert
Weir/Cambridge/IBM]
> 
> -- 
> Andreas J. Guelzow, PhD FTICA
> Professor of Mathematical & Computing Sciences
> Concordia University College of Alberta
> [attachment "signature.asc" deleted by Robert Weir/Cambridge/IBM]