Systems, Man & Cybernetics - April 2016 - 20

ranges. Figure 4 shows two complementary views of the
internal details of the context optimizer. We first examine
the basic problem: that of selecting the appropriate set of
sensors and their settings, given a single context to estimate. We will then look at the more complex problem of
simultaneously estimating multiple contexts.
Single Context Optimization
Given a single context measure, the goal is to choose a
subset i of sensors (and their tolerance ranges) to infer
that context measure with a QoINF value that is at least
equal to the application-specified minimum required
fidelity at a minimum communication overhead. If the
subset, i, of sensors is predefined, then determining the
best tolerance ranges (q i values) is a straightforward
Lagrangian optimization problem. Accordingly, the challenge here is to determine, in the first place, which i to
use. Clearly, one solution is to iterate through all possible combinations of available sensors. However, as sensors become increasingly ubiquitous in our targeted
smart-assisted living environments, such an approach is
excessively computationally expensive. A heuristic
search can drastically reduce the computational cost by
performing an intelligent exploration of the possibilities.
Our proposed heuristic is based on the observation that
the additional cost in adding another sensor to the set i is
dependent on the sensor's hop count from the context estimator and the sensor's sensitivity factors [the h and o
terms in (1)]. Specifically, the algorithm favors sensors with

lower hop counts (indicating a small update cost) and lower
sensitivity factors (indicating a smaller degradation in
QoINF with increasing tolerance ranges) [12]. The heuristic
algorithm first sorts the available sensors based on their
hop counts and sensitivity factors. It then incrementally
considers larger sets of sensors, starting with the singleton
set of the first sensor in the list. The algorithm computes
the tolerance ranges (for each individual member of the set)
needed to ensure that the application-specified QoINF
bound is satisfied and then computes the transmission cost
associated with using those sensors with the corresponding tolerance ranges. If the QoINF requirement is not
achievable with the considered set, the cost is set to 3 .
The algorithm then compares this cost to the cost calculated in the previous round. If the cost has decreased, the
algorithm continues its iterative exploration by growing i.
If the cost has increased, the set computed in the previous
round (and its associated tolerance ranges) is assumed to
be the preferred solution [11].

Multicontext Optimization
To address our eventual vision of a smart matchmaking
service that lets numerous health-care-related applications
and services make the best possible concurrent use of an
underlying substrate of multimodal sensors, we must
extend the algorithm to consider the optimization of multiple distinct contexts [14]. As a tangible illustration of this
scenario, consider again a smart-home assisted-living
deployment scenario depicted in Figure 1, with several sensors: [blood pressure (BP), ECG, passive
infrared sensor (PIR), force-sensitive re sistor (FSR), accelerometer, ultrasonic,
1) Find Hop Count and Tolerance
electromyography (EMG), motion, light,
1) Define Cost Measure
Range for Each Sensor
etc.]. Some of these (i.e., motion, light, PIR,
FSR, and ultrasonic) are embedded in the
2) Specify Application's
environment, and some (i.e., BP, ECG,
QoINF Requirements
2) Specify Minimum QoINF
accelerometer, and EMG) are worn on the
body. Multiple applications, like vital-signs
3) Formulate the Min-Cost
monitoring, fall monitoring, and wellness
QoINF-Aware Problem
management, execute simultaneously
3) Solve Lagrangian
Optimization Problem
using these sensors and require different
4) Solve the Optimization
context attributes at different levels of
Problem
accuracy. For example, the fall-monitoring
4) Compute Ideal
application may require a person's moveTolerance Ranges
ment context to be inferred using BP, FSR,
5) Determine Parameters
of the Sensing Task
and accelerometer sensors, while for the
wellness-management application, context
5) Compute the Cost
describing a person's sleeping state with
6) Compute the Cost
required accuracy can be achieved jointly
by accelerometer, PIR, and ultrasonic sensors. In this simple example, all of the con6) Return the Optimal
7) Apply Search Heuristic
Sensor Set
texts required by different applications
can be satisfied by using only the BP, FSR,
(a)
(b)
and accelerometer sensors (with the
required accuracy and imposed tolerance
Figure 4. Context optimization in a QoInF-aware architecture.
ranges); the other sensors (ECG, PIR, and
(a) a generic view and (b) a parametric view.
20

IEEE SyStEmS, man, & CybErnEtICS magazInE A pri l 2016

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