Systems, Man & Cybernetics - April 2016 - 22

100

2,000
QoINF Accuracy

Reporting Frequency

1,500

1,250

1,000

750

500
250

30
0 10 20 30 40 50 60 70 80 90 100

QoINF (%)

Reporting Frequency

1,750

Tolerance Range (q) in Degrees
Figure 5. Communication overhead and accuracy

tradeoffs.

tradeoff for the motion sensors we used in our experimental deployment.
◆ Using multiple sensor types to jointly infer a single
context metric provides a clear benefit, and our basic
and extended heuristic algorithms take advantage of
this benefit of joint sensing.
A Performance of the Range-Heuristic
Our range-based heuristic can achieve application-specified quality and reduce network resource usage substantially. We compare our range-based heuristic algorithm
with the naïve heuristic and brute-force search. Based on
the derived sensitivity factors, we sort all of the sensors
and generate the following sorted lists: 0 1 = {Shimmer
Accel, SunSPOT Accel, Shimmer Gyro} for context sitting;
0 2 = {SunSPOT Accel, Shimmer Gyro, Shimmer Accel} for

context walking; and 0 3 = {Shimmer Gyro, Shimmer
Accel, SunSPOT Accel} for context running. We use each
approach to compute the optimal sensor set (it ) and asso\(i)
ciated tolerance ranges Q (it ) that minimize the COST
for a target QoINF. We also use the range-heuristic to compute the q values for a target QoINFmin and then use those
q values to determine the achievable QoINF.
Figure 6 plots the minimal cost associated with the
three search methods to determine the optimal subset of
sensors and their tolerance ranges for the first context
state considered, sitting in our case. In this example, the
range-based heuristic and heuristic perform exactly as the
brute force in finding the optimal sensor subset with minimum cost. Figure 7 compares the performance of these
three algorithms for the context walking. The range-based
heuristic performs better than the heuristic, and it performs close to brute force. Similarly, Figure 8 plots the performance for the running context, where again the
range-based heuristic algorithm outperforms the naïve
heuristic. Due to the simple set theoretic addition of sensors from one context to another (without examining the
existing sensor set's satisfiability for the new context) in
the heuristic algorithm, we observe that first just the
Shimmer accelerometer has been selected for the sitting
context; then for walking, both the Shimmer and SunSPOT
accelerometers have been selected; and then for running,
all three available sensors have been chosen. In the rangebased heuristic, only the Shimmer accelerometer is selected for all the contexts at the minimal cost by tightening
the tolerance range.
We also evaluate our range-based heuristic's ability
to attain the application's desired QoINF. First, we calculate the tolerance ranges for the chosen optimal

0.55

16
Range Heuristic Sitting
Heuristic Sitting
Brute-Force Sitting

0.5

12
Minimal Cost (θ )

Minimal Cost (θ )

0.45
0.4
0.35
0.3

8
6

0.2

0.3

0.4

0.5 0.6
QoINFmin

0.7

0.8

0.9

Figure 6. a range heuristic, heuristic, and brute-force

minimal cost comparison for sitting.

0.25
0.2
0.1

Range Heuristic Walking
Heuristic Walking
Brute-Force Walking

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

0
0.1

0.2

0.3

0.4

0.5 0.6
QoINFmin

0.7

0.8

0.9

Figure 7. a range heuristic, heuristic, and brute-force

minimal cost comparison for walking.

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