Database Reference
In-Depth Information
between different events, which goes beyond observing mere correlations
or association rules.
Besides the aforementioned extensions, significant opportunities are
present for landmark results that define new areas. In particular, cur-
rent work broadly adopts one of two distinct philosophies; either (i)
eliminating errors by design, or (ii) troubleshooting them as they occur.
Present literature on sensor network debugging is fragmented along the
above boundary into
pre-run-time
and
run-time
solutions. Pre-run-time
solutions are motivated by the potentially great financial and safety cost
of run-time errors, whereas run-time solutions are sought due to the
scalability challenges and high cost of reliably proving correctness in ad-
vance. For a great category of embedded and networked sensing systems,
both solutions are of value, as they complement each other's limitations.
Hence, an interesting area of investigation lies at the intersection be-
tween the two. It would be of great interest to understand, for example,
how run-time data mining solutions can help build increasingly more
reliable and complete models of subsystems, suitable for formal reason-
ing about correctness when these subsystems are integrated into other
systems. It is also interesting to understand how an incomplete model
of a system (e.g., a detailed model of only some of the components) can
be integrated with data mining solutions to significantly improve the
scalability and accuracy of root-cause diagnosis.
A different open area lies in the topic of learning from experience.
When human experts troubleshoot systems, they come with significant
background and experience inherited from other systems that helps them
identify new problems quickly. How to develop a system representation
model then, such that troubleshooting tools can learn from their own ex-
perience and decide which parts of a learned model remain applicable in
a new context? How to exploit such past experience to substantially im-
prove the scalability and diagnostic accuracy of debugging tools? While
significant work has been done that demonstrates instances of such trans-
fer learning, general solutions that apply experiences across broad cate-
gories of networked sensing installations remain to be found. Note that,
this transfer learning problem, in the context of networked sensing, is
more challenging than its counterpart in general distributed computing.
Many different distributed computing systems are built on the same op-
erating system abstractions, such as threads, processes, memory, system
calls, and synchronization primitives, making it easier to transfer knowl-
edge across troubleshooting experiences. For example, the symptoms
of a deadlock are similar across different software systems. In contrast,
many embedded system installations are unique. They use different
sensors and actuators in different types of physical environments, giving
