Re: [Geopriv] PIDF-LO uncertainty and confidence

[Martin Thomson <[e-mail]>]

On 22 March 2013 06:13, Brian Rosen <[e-mail]> wrote:
> I would like to see this work finish.

As would I.

> I will say that I've been uncomfortable with the position (snap) taken by
> Martin in regard to uncertainty.

I've never had anyone try to explain this discomfort, thanks for trying.


> Basically, the way that he proposes, and in fact the only way we currently
> have, to express a point with uncertainty is to express the uncertainty
> area.
>
> That loses information.

Information loss is inevitable.  The information that the location
server has is not typically usable.  Hence the reduction to an
uncertainty region.

> You start with a  point with uncertainty,

Oh.  That's not actually true.  It's a common misconception, but I
didn't realize that this was the source of the contention.  It is true
in a subset of cases, but more often you reduce a series of
measurements to an uncertainty region without ever having any single
point.

Even if there was a way to express the location of the point that had
the highest probability of being the "correct" point, assigning
special significance to that point is actually dangerous, even if it
was the one point that a location server started with.

That said, the uncertainty draft does describe a way of picking a
point, and an explanation of the consequences of doing so.  That's
probably the most important function the draft provides: explaining
how to extract usable information from what you get.

> I think it would be better to send the original data, and also express the
> uncertainty.

I don't know if you recall Christian Hoene's draft and presentation on
providing richer location data.  It didn't get a lot of interest.  It
did exactly that:
  http://tools.ietf.org/html/draft-hoene-geopriv-bli

This describes how to describe residuals out of least-squares
estimation, or describe a probability density function numerically
(i.e., as a map of area to estimated probability).  It's complicated,
but very close to what a location (information) server actually has
when performing location determination.

The main concern with this sort of approach is not that the original
data is not present, but that there is too much of it and it's not
usable.

> But I think it would be better to be able to express the original
> measurement, and express the uncertainty independently.

That's what draft-ietf-geopriv-held-measurements does.  Unfortunately,
not everyone knows what to do with that information.  It comes back to
usability.  Now, we've made concessions to existing practice in
providing uncertainty shapes that convey more information than
point+scalar uncertainty, but that has a usability trade-off, which
this draft is intended to address.