Systems, Man & Cybernetics - April 2016 - 23

optimization of multiple contexts
subset of sensors at m in ima l
over a common substrate of sencost. Then with those specified
Much of the work on
sors can provide significant savtolerance ranges and the deteri ngs i n energ y overheads for
mined sensor set, we r un our
utility-based context
representative health and wellemulation on the a lready colmodels faces the
ness applications. Accordingly,
lected data traces to determine
practical difficulty
we believe that the technica l
the empirically achieved accuracommunity should explore this
cy of the a lgor ith m. Figure 9
of computing useful
approach fur ther. Our experiplots the QoINF achieved by the
utility functions.
ence with the design and develrange-based heuristic algorithm
opment of this framework has
for t he contex t r u n n i ng. T he
also left us with several insights
range-based heuristic performs
and open questions.
well at no more than 10% lower
than the target QoINF. Nevertheless, we do notice that
◆ What is the right QoINF function for a given
our range-based heuristic does not perform well in
context measure? One of the main challenges in the
achieving target QoINF accuracy for the other two
application of our framework is establishing approcontext states. We believe this is a result of the large
priate QoINF functions for context variables. Much of
approximation in our cur ve-fitting approach. This
the work on utility-based context models faces the
incurs errors in determining the sensitivity factors
practical difficulty of computing useful utility funcand therefore introduces a larger deviation in the q
tions [4]. We have used our inverse-exponential model
values, which ultimately affects the attainable QoINF
and employed regression techniques on training data
accuracy of the range-heuristic algorithm with respect
to derive the parameters for this model [12]. In reality,
to the target QoINF metric. Adding more sensors to
the functional relationship may be not only different
the selection process (as is likely in future pervasive
(for instance, we already know that the qo inf () funccomputing scenarios) would be expected to help limit
tion can be discontinuous) but also deployment spethis negative impact.
cific (e.g., the correlation between a specific user's
movements and motion sensor data may vary signifiInsights and Challenges
cantly based on individual behavioral characteristics
Our initial work with this framework and its implemenor the layout of an assisted-living facility), and a sepatation provides enough evidence that our suggested
rate training phase may be impractical. In such situaapproach of a) building formal functional models to
tions, we need to explore a more continuous, adaptive
characterize the relationship between context attrilearning framework, where the system dynamically
butes and individual sensors and b) applying joint
learns the relationship be tween different sensor

0.9

Range Heuristic Running
Heuristic Running
Brute-Force Running

0.85

Range Heuristic QoINF Accuracy

QoINF Accuracy

Minimal Cost (θ )

0.8
15

0.75
0.7
0.65
0.6

5
0.55
0
0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

QoINFmin
Figure 8. a range heuristic, heuristic, and brute-force

minimal cost comparison for running.

0.5
0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95
Target QoINFmin
Figure 9. a range-based heuristic on achieving

target QoInF for running.

Ap ri l 2016

IEEE SyStEmS, man, & CybErnEtICS magazInE

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