Systems, Man & Cybernetics - April 2016 - 21

without altering the set of actiultrasonic) may be then turned off
vated sensors, simply by tightento conserve energy.
One of the main
ing the tolerance ranges of the
The preceding scenario can be
challenges in the
current set of selected sensors.
expressed as a multiattribute
We have thus far assumed that,
optimization problem, whose goal
application of
for a given context, the user is
is to achieve the required QoINF
our framework
only in one context state at a
of multiple applications, while
is establishing
time, i.e., either the user is in the
simultaneously minimizing the
sitting state, in the walking state,
total communication overhead
appropriate QoINF
or in the running state. There are,
[12]. The extended heuristic algofunctions for context
however, other scenarios like
rithm for solving the multicontext
watching telev ision (TV ) and
problem considers the added
variables.
speaking on the phone, which
dimension of the problem; specifimay happen concurrently. Such
cally, the set i that is best for a
concurrent context states can
particular context metric in isolaalso be determined using our model. As shown in our
tion may no longer be ideal when considering contexts
model, the minimum QoINF value and sensitivity factors
jointly. The heuristic algorithm still favors sensors with
for these multiple context states will be fundamentally
lower hop counts and lower sensitivity factors, but a sendifferent. For example, consider that we have one acoussor's sensitivity factors are dependent on the particular
tic sensor for detecting the watching TV context state
context being inferred. As a result, if we have L different
and one microphone sensor for recognizing the speaking
contexts to jointly estimate, we have L sorted lists of the
on the phone context state. The operating analytics (tolavailable sensors. Our goal is to satisfy all L contexts at
erance range, etc.) of these two sensors can be computed
the same time; our algorithm considers them sequentialby our proposed model while still maintaining the underly. When only the first context (C1) is considered, its sortlying objective of sharing sensor data streams to
ed list is used, and i is constructed exactly as in the
improve the accuracy and minimize the network cost.
single context case. When the algorithm moves on to the
second context (C 2), it first determines whether the
Evaluation
existing i is sufficient for estimating C2. If not, the algoTo illustrate the promise of this formal model-based
rithm adds new sensors using C2's sorted list. As it does
approach, we experimented with a laboratory-based
so, it also tests whether any sensors that were added to
deployment in which individuals were monitored with
support C 1 have become redundant; if so, they are
body-worn sensors taking readings from a motion sensor
removed. The algorithm continues this process incre(an accelerometer), a light sensor, and a temperature senmentally until it has considered all L contexts.
sor. We have performed experiments with SunSPOT
(www.sunspotworld.com) and Shimmer sensors (http://
A Range-Based Sensor Selection
www.shimmer-research.com/). Specifically, we have used
for Multiple Contexts
a three-axis accelerometer, a light, and an embedded
Here, we discuss an enhanced version of the previous
external gyro sensor on the SunSPOT platform and a
heuristic algorithm that, for each additional context,
three-axis accelerometer and gyro sensor on the Shimmer
tries to compare the total cost from the following two
platform. This initial deployment gives important insights
approaches: 1) using the current set of sensors and
in the nature of context inference and the use of awaredetermining if a modification of the tolerance ranges of
ness of quality to direct the acquisition tasks [11], [12]. Our
this current set is enough to satisfy the QoINF requireexperimental data and results can be summarized via the
ment of the additional context metric or 2) adding an
following key observations.
additional sensor to the set of sensors and seeing what
tolerance ranges this modified set must have to satisfy
◆ A clear relationship between QoINF and tolerance
all the QoINF requirements of the contexts considered
range exists, but this relationship is neither linear
thus far. After computing the costs of each approach,
nor continuous; for some data types, the quality of
this second heuristic selects the one that both satisfies
(context) inference possible using the data type can
the QoINF requirements of all of the considered condrop precipitously with just a small change in tolertexts and has the lowest cost. This is in contrast to the
ance range.
approach in the previous algorithm, where the compari◆ The expected relationship between cost and tolerance
son was made only between adding a new sensor and
range exists: raising the tolerance range decreases the
the cost incurred by the current set of sensors (with
cost. Taken with the previous observation, it is possitheir tolerance ranges unmodified). In other words, the
ble to exploit the tradeoff between quality and cost by
previous approach did not explore the option that one
tinkering with the tolerance ranges, and this tinkering
could satisfy the QoINF of the additional contexts,
is specific to particular data types. Figure 5 shows this
Ap ri l 2016

IEEE SyStEmS, man, & CybErnEtICS magazInE

http://www.sunspotworld.com http://http:// http://www.shimmer-research.com/

Table of Contents for the Digital Edition of Systems, Man & Cybernetics - April 2016