Database Reference
In-Depth Information
lar mathematical or logical invariant. For instance, if a large number of
cars were involved and many unit incompatibilities were present, rather
than returning all combinations of cars that report measurements in
conflicting units as discriminative, it would have been good to return a
single rule of the form
A
=
B (unit)
, indi-
cating that the discriminative subgraph features two arcs (between two
sources and a common sink) feeding data of different units. The rule
could then have support that enumerates all found instances of incom-
patibility satisfying the above condition. Identifying such symbolic rules
could significantly improve the readability of diagnostic results.
An attempt to handle this problem is described in recent literature [47],
defining a
symbolic pattern
as one where all or a subset of the absolute
values of event attributes within the discriminative pattern (that rep-
resent a potential bug triggering condition) are replaced with
symbols
to generalize the pattern. In this case, the logged execution events can
include any operations performed at runtime such as message transmis-
sion, message reception, and writing to flash storage. Each recorded
event can have multiple attributes. The symbolic pattern extraction
algorithm first generates discriminative patterns then generalizes them
by mining for “relationships” across those patterns. A new scheme is
presented for counting the support for individual patterns which greatly
enhances the chances of identifying “infrequent” events that are cor-
related with failure. The resulting patterns are ranked, such that most
informative patterns are presented first. The authors demonstrated that
the approach returned significantly fewer yet more informative patterns
compared previous discriminative mining solutions that had no symbolic
generalization capability.
→
→
C,B
C: A (unit)
4. Other Sensor Network Debugging Work
There has been a substantial amount of work on automating sensor
network debugging, attributed to the inherent diculty of addressing
this problem manually. While the current chapter surveyed approaches
based on data mining algorithms, the list of different debugging tools
and techniques does not stop there.
Aside from techniques inspired by data mining, work on sensor net-
work debugging started with the development of appropriate
laboratory
testbeds
, such as Motelab [94], Kansei [26], and Emstar [32], that facili-
tate sensor software testing by providing the convenience of a controlled
experimental environment. Early work also included tools that improved
the visualization of network statistics. One of the first examples in that
area is SNMS [88]. It constitutes a sensor network management service
