Database Reference
In-Depth Information
Problem 2.1 (Model-Driven Data Acquisition)
Given a sensor
network, and a sink that needs to collect all the sensed values within
of the real value with confidence (probability) at least
1
δ
, design a
data collection protocol such that the energy used by the sensor network
is minimized.
−
In order to solve this problem, we need to decide on the probabilistic
models to use for approximating the distributions of the sensed values,
and also on the communication strategies among the sensors and the
sink. Both these aspects of the problem are addressed by the studies
that we discuss in the next paragraphs.
2.1.1 Proposed Techniques.
The BBQ system [32] proposes
sensor data acquisition techniques based on time-varying multivariate
Gaussian probabilistic models, but other models can alternatively be
used, such as probabilistic graphical models [31]. Using the above ap-
proach, the produced models capture correlations both among sensed
values from the same sensor across time, and among different sensors
across space. We note that the above approach requires some knowledge
of the special characteristics of the data distribution, such as periodic
drifts, which should be encoded in the space of models considered. This
means that some minimum amount of domain knowledge is required, in
order to make effective use of these techniques.
A similar framework for modeling sensor network data is proposed
by Guestrin et al. [45]. The goal is for groups of nodes in the net-
work to collaborate in order to fit a global function to each of their
local measurements. This approach employs kernel linear regression in
order to model the sensed values, by capturing spatio-temporal correla-
tions. Once again, we observe that this is a parametric approximation
technique, and as such, requires the user to make an assumption about
the number of estimators required to fit the data. Moreover, there is a
need for a training phase (where the models are built, evaluated, and
adjusted), which in practice can be rather lengthy and expensive.
Even though the domain knowledge requirement that the above tech-
niques have may be prohibitive for some applications, we note that a
large number of applications (where the measured phenomena are known
or understood, or when a domain expert is available) can still benefit
from such techniques.
2.2 Data-Driven Data Acquisition
The model-driven approach described earlier can lead to significant
energy savings for the data acquisition task. However, by the nature
