Re: [apps-discuss] Draft on representing values of sensors in JSON (draft-jennings-senml)

["Thomson, Martin" <[e-mail]>]

Hi John, Cullen,

On 2010-10-26 at 04:54:59, John C Klensin wrote:
> Confidence expressed as a single probability and uncertainty
> expressed as a single value inherently assume that the
> distribution of the measure is central, symmetric and, for most
> purposes, unimodal.  While it does not assume normality, one
> might end up making that assumption to interpret the numbers.

A receiver makes such assumptions either because they know they can (out of band information), because they are ignorant of the possibility, or because they know the consequences and are prepared to accept them.

This format could convey a wealth of information on the observed or theoretical probability distribution of the measurand.  That's probably overkill in these sorts of use cases.  The unc and c parameters are there when cheap is good enough.

In cases where it's important, a sensor can report discrete observations.

> The latter would imply that "c" and "unc" would need to be used together or not at all.  

Yes, there is a pairing.  Usually it goes like this:  c has a default value (0.95); unc doesn't, and is assumed to be unknown in its absence.

[I just noticed that the updated draft says that confidence should be assumed to be "approximately 1".  Better to say "exactly 0.95".  1 is impossible; 0.95 is plenty high, and plausible.]

> In addition, if you are going to deal with "integrated sums" and
> their definition, I suspect that you should be prepared to deal
> with moving averages and autocorrelation functions: even if few
> sensors produce those today, they can be extremely valuable
> statistically for measurements of some types of phenomena,
> especially when deviations from some norm are more significant
> than the norm itself or when measurements are to be used
> predictively.

Moving averages and the like were certainly considered.  A sensor that runs a Kalman or particle filter is not outside the realms of possibility.
 
> I don't think the fairly subtle statistical considerations I've
> pointed to above ought to need treatment in the present
> document.  Indeed, although I could provide some references to
> systems that tried, to do so would, IMO, take you down a rat
> hole.  But I think the document should be very clear about
> expansion paths ...

Sounds reasonable to me.  Rat holes swallow a lot of good work.

--Martin