The description "The Smart Sensor Web" concentrates
on 'plumbing', that is, just getting data. The
description is too light on the rest that is needed
for real usefulness.
The description appears to fall into an old trap,
the promise that with all the data we have
'everything'. Yes, getting the data is usually
necessary. However, the data alone is rarely
sufficient and, thus, not yet 'everything'. So, we
also need (1) dictionary of the data, that is, what
data is where (e.g., what Google does for the Web),
(2) descriptions of the data, e.g., as in XML, the
older OSI CMIS/P ideas, and/or just natural language
(the sensor web data will usually not be
self-documenting -- there is a challenge here not
faced by the Web), (3) what we are going to do with
the data (we can't expect just to have humans read
it), and (4) what the real and valuable applications
are (not the news and entertainment of the Web).
The description did mention "intelligent" and
"information layer", but it is here that the crucial
issues are for power and value; thus, we need much
more than just the simple mention.
Broadly we can compare with the Web -- TCP/IP with
HTTP and HTML: The Web mostly presumes that the
server is sending to a PC with a human reading a
screen. So, the Web got to exploit the ability of
humans to read screens.
For sensor webs to yield valuable results, we need
some powerful automation of the data, need to
replace the human reading a screen. There is value
here but also challenges.
A guess: Too soon, we will want more than just
'sensors'. We will also want 'transducers' that let
us 'control'. Also, we will need security, etc.
Depending on context, difficulty of 'validation' can vary
widely from quite easy to far too difficult.
Here is an outline for how to make 'validation' easy:
Generally we are dealing with systems that cannot be predicted
exactly in advance. Here our usual best approach is to be
'probabilistic', and for this a good first step is to look
for cases of 'independence'. While there are statistical
'tests' for independence, usually we believe in independence
based on what we know about the system broadly and mostly just
intuitively.
Independence is powerful: We get to use the classical
convergence results: law of large numbers, central limit
theorem, etc. And independence brings some 'symmetry' results
that can be powerful.
Some results weaker than independence can also be powerful:
Often an ergodic assumption can be powerful enough and easy
enough to believe in.
Many of the classical statistical techniques were from the
days of small amounts of data and very poor computing. With
the large amounts of data and current computing, we can
consider newer techniques.
Exploiting such assumptions, that actually we got mostly just
from broad understanding and intuitive means, we may be able
to get all we need for good 'multivariate statistical quality
control'.
In this way, 'validation' can be easy.
In the post
kcelery (410487) on Monday October 06, @03:07AM (#7141869)
is mention of false alarm rate. Right, it is important and
for the reason mentioned: False alarms can be costly.
Really, missed detections can be more costly.
These two, false alarm rate and detection rate, are the main
content of 'quality' in detection and monitoring.
Generally, between these two, false alarm rate is the easier
to control, i.e., in real situations we usually do not have
enough information to apply the Neyman-Pearson lemma.
Slashdot Mirror
The description "The Smart Sensor Web" concentrates on 'plumbing', that is, just getting data. The description is too light on the rest that is needed for real usefulness.
The description appears to fall into an old trap, the promise that with all the data we have 'everything'. Yes, getting the data is usually necessary. However, the data alone is rarely sufficient and, thus, not yet 'everything'. So, we also need (1) dictionary of the data, that is, what data is where (e.g., what Google does for the Web), (2) descriptions of the data, e.g., as in XML, the older OSI CMIS/P ideas, and/or just natural language (the sensor web data will usually not be self-documenting -- there is a challenge here not faced by the Web), (3) what we are going to do with the data (we can't expect just to have humans read it), and (4) what the real and valuable applications are (not the news and entertainment of the Web).
The description did mention "intelligent" and "information layer", but it is here that the crucial issues are for power and value; thus, we need much more than just the simple mention.
Broadly we can compare with the Web -- TCP/IP with HTTP and HTML: The Web mostly presumes that the server is sending to a PC with a human reading a screen. So, the Web got to exploit the ability of humans to read screens.
For sensor webs to yield valuable results, we need some powerful automation of the data, need to replace the human reading a screen. There is value here but also challenges.
A guess: Too soon, we will want more than just 'sensors'. We will also want 'transducers' that let us 'control'. Also, we will need security, etc.
Depending on context, difficulty of 'validation' can vary widely from quite easy to far too difficult.
Here is an outline for how to make 'validation' easy:
Generally we are dealing with systems that cannot be predicted exactly in advance. Here our usual best approach is to be 'probabilistic', and for this a good first step is to look for cases of 'independence'. While there are statistical 'tests' for independence, usually we believe in independence based on what we know about the system broadly and mostly just intuitively.
Independence is powerful: We get to use the classical convergence results: law of large numbers, central limit theorem, etc. And independence brings some 'symmetry' results that can be powerful.
Some results weaker than independence can also be powerful: Often an ergodic assumption can be powerful enough and easy enough to believe in.
Many of the classical statistical techniques were from the days of small amounts of data and very poor computing. With the large amounts of data and current computing, we can consider newer techniques.
Exploiting such assumptions, that actually we got mostly just from broad understanding and intuitive means, we may be able to get all we need for good 'multivariate statistical quality control'.
In this way, 'validation' can be easy.
In the post
is mention of false alarm rate. Right, it is important and for the reason mentioned: False alarms can be costly.Really, missed detections can be more costly.
These two, false alarm rate and detection rate, are the main content of 'quality' in detection and monitoring.
Generally, between these two, false alarm rate is the easier to control, i.e., in real situations we usually do not have enough information to apply the Neyman-Pearson lemma.